Blessed thread of frenship

File: FLUX_00009_.png (911 KB, 896x1152)

911 KB PNG

bleh

>mfw

File: 00021-3327940258.png (2.13 MB, 1432x1432)

2.13 MB PNG

>>101850830
>>101850957
It'll take ages for Flux to catch up to Pony, It's pretty much everything we can want in a model right now baring a few issues.
But by god i want more improvement, even if it means Pony jumps to Flux for a base and i cant run it until i get my Ai system kek

>>101851081
Pony is pretty disappointing as soon as you venture outside of the material it was created for

Cursed thread of bullshit

File: FLUX_00086_.png (1.2 MB, 896x1152)

1.2 MB PNG

flux likes making geeks asian
racist

File: Capture.jpg (113 KB, 2752x748)

113 KB JPG

>>101850925
>And this >>101849076 thing gives me the same shit.
that's normal for that one, that TonemapNoise shit needs to be deactivated when the CFG returns back to 1, a bit like DynamicThreshold did a few days ago

What you need to do is this:
1) Go on custom_nodes\ComfyUI_experiments
2) Open the three python files (sampler_rescalecfg.py + sampler_tonemap.py + sampler_tonemap_rescalecfg.py)
3) Add this part of the code to make it work with AdaptiveGuidance:

if cond_scale == 1.0:
return cond

Someone should make a PR on his repo so that it will work for everyone when updated
https://github.com/comfyanonymous/ComfyUI_experiments

>>101851184
Pony is made specifically for porn, why would you ever use it for anything else when there's a dozen of good normal SDXL models for that?

File: 00005-natvisNaturalVision_v10.png (1.21 MB, 832x1216)

1.21 MB PNG

>>101851081
for realism pony only needs a good merge with natvis I feel like. I only lacks prompt adherence compared to pony.

>>101851315
Some of the best 1.5 merges I have were made for porn but that didn't stop them from being able to do just about anything.
That said, I recognize that is is 10 times shittier to train XL than 1.5 and I can't really expect too much (It's still fucked up pony gives you a goth girl if you type in shoggoth)

File: result.jpg (283 KB, 800x836)

283 KB JPG

I am trying these Loras as they come out and getting disappointed. This is with the Impressionism Lora.
Same prompt and seed. Important details is the blue dome and the castle above the town on a crag.
Flux gets things right and ignores the style completely.
Then I use the result of flux to img2img with SDXL. (Dreamshaper turbo). It doesn't score too well in style, but at least it is trying, there goes the details.
Flux+Lora: Aesthetics down the drain, but at least looks in the park of Impressionism. To me it looks amateurish. The details are lost for the most part, there is nothing that you can clearly identify.
That output I reprocess with the SDXL model, and the style reverts to the same style as in the first case that doesn't look Impressionist enough.

>>101851451
Flux is shovelware, dude. It's a gimped model that is aggressively censored dumped on the open source community that would be a bitch to unfuck. Companies do this as a way to make it look like they're not anti-open source when they in fact are.

Back at it again in the Popular Threads section.

File: ComfyUI_00016_.jpg (866 KB, 2304x1792)

866 KB JPG

>>101851493
>t. vramlet

>>101851493
complaining about free shit is mental illness

File: 2024-08-12_021715_seed64_(...).png (1.91 MB, 1536x864)

1.91 MB PNG

I let the thing autogen for a while and got some ones I liked. I'm not going to post all of them so here's a zip if you want to see them (also if you want the prompts/workflow). Warning: some of them are pretty spoopy :)
https://files.catbox.moe/zyp153.zip

>>101851451
>Flux+Lora: Aesthetics down the drain, but at least looks in the park of Impressionism. To me it looks amateurish. The details are lost for the most part, there is nothing that you can clearly identify.
It's the only one that looks actually impressionist. But it has lost all the concepts from the original model.

File: 2024-08-12_031120_seed147(...).png (2.06 MB, 1536x864)

2.06 MB PNG

>>101851722

flux dev vagene lora eta?

File: ventus ok i believe you.gif (1.17 MB, 480x270)

1.17 MB GIF

>>101851801
ok i believe you pedanon

>>101851876
>but that's not allowed under flux-dev license
really?? I wasn't aware of such a jewish license. Man i really do need to catch up on the news about Flux.
Im willing to put my hopes in robin hood scenarios like youre outlining, given how many we've had over the past year for basically the entire A.I world and our best moments came from leaks/dudes just propelling us entirely off solo work.

File: image-grid.jpg (1.27 MB, 3076x3076)

1.27 MB JPG

>>101851388
A lot of the SD1.5 anime models were a ton of fun because 1.5 has the most artists and mediums baked into it and I love the whole textured look of it. It excels at being abstract and stylized even though what it creates is kinda poorly defined (LOTS of mangled hands and what not). I really have trouble getting anything like these pics for example on newer models, they come out too clean and defined though Loras can help a bit.

File: FLUX_00004_.png (1.65 MB, 1024x1024)

1.65 MB PNG

it's like theme park, as in a shittier RCT

File: ComfyUI_01134_.png (263 KB, 1024x1024)

263 KB PNG

>>101851315
>>101851184
Pony has a great understanding of human body and poses. Other models can't comprehend even basic expressions and limb positioning. It's very useful outside of porn.

File: flux_00147_.png (1.76 MB, 1344x768)

1.76 MB PNG

just dont look too close

>>101851979
nice

File: ComfyUI_temp_lugxa_00114_.png (1.87 MB, 1184x864)

1.87 MB PNG

>>101851732
As a proof of concept it works. The suggestion is strength 1 but in my oppinion you lose too much detail, to the point of that you don't know what you are looking at; that I think isn't Impressionism but Abstract Expressionism. Using a lower strength works for me.

File: ComfyUI_00936_ (1).png (413 KB, 1024x1024)

413 KB PNG

>>101851901
Shnell version doesn't have it, so at the very least we can always have a smaller model. Still can be finetuned to perfection.

File: ComfyUI_Flux_125.png (1.29 MB, 1344x768)

1.29 MB PNG

>>101850975
>>101851142
>>101851302
Nope, still doesn't work no matter what. Updated, deleted, installed again, edited, re-edited files, nothing works. There was even an update to AdaptiveGuider node half an hour ago, and still nothing. Maybe it's a torch/cuda issue on my side, I don't know, will try updating them later. Meanwhile I'll just stick around with AutomaticCFG warp node which also works with cfg > 1 and adaptive guidance.
On the side note, SamplerDPMPP_2M_SDE with heun solver and ays+ as a scheduler looks pretty slick at 25+ steps

>>101852170
did you use some workflow or you made it by yourself? maybe that could help

File: 1708700932491506.png (2 MB, 1024x1024)

2 MB PNG

this is thumbelina apparently.
thank fuck these models are local, i can already see all image gen going the route of vidya and only producing fugly niggers.

File: Capture.jpg (170 KB, 1265x1777)

170 KB JPG

>>101852170
>On the side note, SamplerDPMPP_2M_SDE with heun solver and ays+ as a scheduler
what? I don't have a "solver" or a ays+ on ComfyUi

>>101852212
it already did, just the cloud models. We're lucky as FUCK that imagegen didnt just die this year.

Fix xformers please

>>101852226
nah, i have no idea what the story ended with, but jewgle consneeded that its nigger generator was faulty. afaik jewgle's ai was the only one to gen such poz.

>>101852244
nah bing was doing it too, i forget what the other big tech one was as well but they were all and still are doing it.
even Grok has its biases in spite of how much people retardedly think musk is le based

>>101852264
>in spite of how much people retardedly think musk is le based
I have no idea why the AI bros are still praising him, he literally signed that paper that asked the AI developpment to stop for 6 months

>>101852264
>nah bing was doing it too
it is still doing it but was much smarter about it from the start than google

File: 26699d2839587c5fb9bdf5a9c(...).jpg (500 KB, 896x1152)

500 KB JPG

Your cyber waifu.

File: ComfyUI_01013_.jpg (745 KB, 1440x1920)

745 KB JPG

https://civitai.com/images/23937373
Holy fuck that's pretty good, it really looks like MJ niji

File: FLUX_00069_.png (1.18 MB, 896x1152)

1.18 MB PNG

>>101852319
nah

File: ComfyUI_00149_.png (1.78 MB, 1024x1440)

1.78 MB PNG

https://civitai.com/models/639737/emma-watson-lora
The likeness this lora can achieve being trained on only 25 images is quite incredible

File: db84ec8037ecf6c46096f3416(...).jpg (506 KB, 896x1152)

506 KB JPG

>>101852334
It's kinda funny: easier to get flux to objectify robot bitches than it is to do the same for normal women.

>>101852321
Ehh? How are we supposed to know the trigger words for that one?

>>101852348
Ikr, real finetunes of flux will be so fucking easy, if only 20 pictures is enough to add a new concept into it

>>101852328
One thing I'm not happy about with flux is how it's fucked from a weights standpoint (well, Pro™ probably isn't which is why it's API only). This is becoming a sad pattern in AI dev, although it's better than the status quo ante of a few years ago with openAI just controlling everything. Very few are releasing open weights models that actually are just good without any built-in attempts to poison the well.

>>101852334
prompt for this? can you try to get some exposed wiring/mechanisms too?

>>101852217
NTA, ays+ is probably from https://github.com/pamparamm/ComfyUI-ppm

>>101852436
>Very few are releasing open weights models that actually are just good without any built-in attempts to poison the well.
no startup wants to be known as the "child porn" or "deepfake" company. stable diffusion is barely 2 years old, we've come a long way so far. hopefully flux spooks the chinese into making a great open comparable image model too, and then we can ignore its license entirely like how the chinese ignore ours

File: 1697352594211952.png (1.89 MB, 1024x1024)

1.89 MB PNG

>>101852264
sigh, thats fucking depressing.

>>101852455
dunno, I just like a human face for expression on a plastic body
if androids were real I'd like realistic skin but that doesn't translate into images well

A brunette woman with a white plastic android body, pushing a shopping cart through a supermarket aisle.

>>101852486
>white plastic android body
thanks anon, I was trying to get something like that with "robot" and "cyborg" because that's what used to work on juggernaut but it wasnt working with flux

>>101852383
People should try training embeddings for flux: a1111 could do it, I assume forge can probably too? Or maybe you'd have to train the embeddings using T5-xxl on its own. But between clip and t5, there are probably a lot of concepts that are already in the weights, we just don't know the words.

I get the sense that a good portion of Flux's training data was captioned by ChatGPT, so to reverse engineer you could also just ask ChatGPT to write a caption for an image and then see what happens when you put that caption into FLuX (don't ask ChatGPT to write a prompt – just a caption).

File: ComfyUI_00155_.png (1.71 MB, 1024x1440)

1.71 MB PNG

>>101852348
I don't know how it's possible to get this from 25 images, it can do all angles and is very flexible with expressions.

>>101852472
oh ok, thanks anon

>>101852543
it has already seen millions human faces and bodies, it just needs the little bit of extra information that makes one look like Emma Watson.
Better models have better loras.

>>101852543
>I don't know how it's possible to get this from 25 images, it can do all angles and is very flexible with expressions.
Flux already knew who Emma Watson was (even though it wasn't perfect), it just needed a little boost

>>101852543
pretty much >>101852589
styles loras and some unique character loras start falling apart quite quickly

>>101852478
I am sympathetic but at the same time, I don't want open source to become a junkyard where companies habitually pollute the ecosystem with products they have deliberately sabotaged while gaining good PR. Flux is good enough that it is probably worth the trouble of figuring out how to fix it, but if it was any less capable, it wouldn't be IMHO.

>>101852543
Pictures of her are almost assuredly in the data.

File: 1715522006933425.jpg (284 KB, 1103x639)

284 KB JPG

life is pain. pain is life. pain is pain. pain pain pain.
any hope on the horizon, or is this the new normal if i want quality gens?

>>101852634
>gaining good PR
good thing that doesn't happen. google got backlash for black nazis. stability got backlash for a broken SD3-medium. stop dooming about stuff that isnt happening
>Flux is good enough that
if flux wasn't good enough it wouldn't have been released doomer dummy

>>101852321
is there anything other than close ups? any complex actions?

File: FLUX_00095_.png (1004 KB, 896x1152)

1004 KB PNG

I thought I heard that flux doesn't do gore

File: ComfyUI_31382_.png (1.37 MB, 848x1200)

1.37 MB PNG

>korean runes
Oh God please don't train models on Korean and Chinese "manga", gens end up looking just like the real deal.

>>101852816
dios mio terminator 2 lookin ass

>>101852321
How do we use that? He hasn't said anything about what word to trigger

>>101851514
i like dis one

whats with the pedo posting?

>>101852985
got chased out of /b/ somehow

>>101853035
No it's the thread schizo seething over how dead his thread is

>>101852348
I fucking hate her face, I never understood people obsessed with her.

>>101852985
Report and ignore.

>>101853092
She also let a fat pig fuck her for a now dead career
Disgusting bitch

>>101853125
oi Pedanon, who's your favorite HP girl? Are you a LunaCHAD?

>>101853203
oh god dammit you're more based than i was worried you'd be

>wait why aren't you just genning freckled little girls then?

>>101852478
well as it should be obvious by the faggot in this very thread, if you don't censor the model a couple mentally ill people can literally ruin it for everyone, all it takes is those images floating around social media and the outrage mob starts putting pressure on politicians

>Upgrade pytortch to a version that doesn't support xformers
>have made no efforts to fix this
Why do this Illyia?

>>101853340
GOOD DEPENDENCIES MORNING SAR NEW PACKAGES UPDATE READY FOR YOU YES? DOWNLOAD requirements.txt YES?

>>101853361
kek'ed

>>101853352
Doesn't matter, you can literally put pressure on all the rest of the world, the US has the power to take down CivitAI, the US has the power to take down AI-related code on Github. They can make you operate only on the darkweb. Keep in mind, if it's against US law, no more threads here either.

File: file.png (41 KB, 464x445)

41 KB PNG

>>101852217
Meant these on the pic
>>101852190
I loaded another workflow and only then it struck me - I was using the wrong fucking node. There's two Dynamic Thresholding Full nodes, one from the og repo and another one from EasyUse node group, and I was using the second one. I loaded the proper one and of course it immediately works. Fuck me.

File: ComfyUI_31390_.png (1.67 MB, 848x1200)

1.67 MB PNG

Nigga, just ilm daat walk. That waw!

>>101853455
CSAM is illegal so good luck with your first amendment claim. Also your images are enough for a referral to the feds, have fun.

File: flux_upscaled_00092_.png (3.75 MB, 1856x2400)

3.75 MB PNG

>>101853169
nice

Is flux gonna take off when it's more than 10x time to generate single 1024x1024 image compared to SDXL (pony)?
I'd imagine it would be pretty hard to do good prompt research and fast prototyping when generating big comparison grids takes hour(s).

Also flux lacks so many things that you have in SDXL. Massive library of loras, finetunes, mixes, controlnets and just overall massive amounts of information and tutorials. By the time flux catches up to that, surely there are much better base models available.

>>101853538
The point of having more parameters is it's essentially knowledge size. That means you theoretically don't need to rely on Loras. I'll contend that Flux is basically overbaked on a relatively limited knowledgebase. It is very possible with a series of merges across many properly captioned domains you'd have a much more capable model without relying on Loras which rape the weights.

File: ComfyUI_01337_.png (3.56 MB, 1344x1728)

3.56 MB PNG

>>101853538

>Also flux lacks so many things
>By the time flux catches up to that, surely there are much better base models available.

Those base models will also require a massive library of loras, finetunes, mixes, controlnets and just overall massive amounts of information and tutorials, anon.

Better to work with what we've got right now rather than not making any headway and just hoping something better comes along.

File: flux_upscaled_00082_.png (2.8 MB, 2400x1368)

2.8 MB PNG

>>101853538
literally the same thing was said about SDXL when everyone used 1.5. Optimizations, finetunes and better cards will come

>>101853625
We're already further along and the training ecosystem has matured. The first big step will be someone making a trainer that can use multiple 4090s to train larger Loras and/or do a full fine tune.

So you need a 3090 / 4090 / etc to train flux? Is it now time to buy servers?

>>101853538
>Is flux gonna take off when it's more than 10x time to generate single 1024x1024 image compared to SDXL (pony)?
more like 3x on my 3060 and half that with nf4, it's not that bad

>>101852699
Companies absolutely get good PR for seeming to support open source, or else the nickname "ClosedAI" wouldn't exist, retard.

>>101853446
>I loaded another workflow and only then it struck me - I was using the wrong fucking node. There's two Dynamic Thresholding Full nodes, one from the og repo and another one from EasyUse node group, and I was using the second one. I loaded the proper one and of course it immediately works. Fuck me.
kek, I'm glad you figured out the way to fix this shit, have fun with DT anon

>>101853535
ones on the left be like
>i'm next

>>101853538
I've gen'd probably thousands of images with Schnell. I assume any prompt that works for that would work for the full model but I also don't know which prompts that don't work well with Schnell might work okay with the full model.

>>101853612
This is why people need to be training embeddings and not LoRAs.

>>101853800
>This is why people need to be training embeddings and not LoRAs.
what's the difference?

File: Bows.png (3.21 MB, 2432x1092)

3.21 MB PNG

>>101853538

It takes me about 45 seconds to generate a 1024x1024 image on Flux.
Took 10 seconds to make a same sized pic on SDXL
Perfectly usable.

Even in it's base model state Flux mogs the absolute fuck out of XL in so many ways, that you don't need so much of that extra shit to make it good.
We'll need is the same artist and character loras that currently exist for other models, but a lot of things simply aren't necessary due to the incredible text adherence Flux has.

Pic related is Cinenaut, Cyberrealistic and Flux.
Concept of firing a bow is something that models have really hard time understanding. They can't deal with those strings or even the pose itself.
I cherry picked better results for the two XL models out of many generations, but I simply took whatever was the best from the first Flux batch of images.
It's not even comparable how much better it is understanding prompt and concepts. You need either controlnet or photoshop to get the same result out of other models.
Give it 2 months of style loras and enabling porn generation and Flux will have kicked the current model's asses easily, aside from furry models.

>>101853826
embeddings are a just few kilobytes, no changes to the unet or text encoders
you're just optimizing a few "words" (vectors) of the prompt to pull some concept out of the text encoder. works best with things the encoder already knows but even with just a few vectors it can be something novel.

File: ComfyUI_00166_.png (263 KB, 1024x1024)

263 KB PNG

>>101853538
It's 15 seconds for full fp16 flux dev on 4090, so it's basically nothing.

>>101853935
braggart

File: You know.png (2.68 MB, 1741x1387)

2.68 MB PNG

You just...

Flux is too large for what it knows anyway, could be trained at half size easily

>>101853935
Assuming you're not using negative cfg or an advanced workflow. 25 secs for me.

File: file.png (4 KB, 363x48)

4 KB PNG

I want to try out flux but I'm not sure which model to download. I see a bunch of versions. I get the fp8/nf4 part but whats the difference between dev and schnell?

>>101854099
schnell means fast, so it's probably faster

File: ComfyUI_00245_.png (2.33 MB, 1248x1664)

2.33 MB PNG

>>101854099
schnell is a 4-step model, like SDXL-Lightning models. Dev is the 'full' 20-step model. There's a Pro model as well, available only through their webapp

>>101854132
and faster in machine learning usually means worse quality

>>101854132
>>101854136
I see. I hate lightning so I assume I hate schnell. Ty.

>>101854199
goes without saying

Should I buy a car or should I buy RTX 5090 and a bicycle...

>>101854248
work harder and buy a H100

File: 00034-2950443613.png (415 KB, 600x600)

415 KB PNG

Oi mate, did you just post a meme?

How much VRAM do you think the mid range 50 cards will have? 16gb Seems like a good spot, that's what the 4070 ti super has at least

File: ComfyUI_00209_ (1).png (189 KB, 1024x1024)

189 KB PNG

>>101853851
Can someone post the most basic bitch workflow for flux lora? Every workflow I found requires shitton of custom nodes that just won't work for me

>>101854317
1GB less than you need to comfortably fit the best local image gen available at the time

>>101854317
These are just placeholders but I mean ............................................
https://www.techpowerup.com/gpu-specs/geforce-rtx-5090.c4216
https://www.techpowerup.com/gpu-specs/geforce-rtx-5080.c4217

sup gents
I've been using automatic 1111 should I move to forge or something else, what is the state of the art today?
I have been out of the loop for a while now.

>>101854443

Move to Forge, it's way faster and overall better version of 1111.
I made the move when Flux became a thing and it's simply a superior option.

>>101854400
>RTX 5080
>16gb
So 12gb is likely for the 5070 then?

>tfw my fetish is thicc downies
>every image generator struggles to produce faces that look remotely like real downies
Fuck my life. Maybe I have to get off my lazy ass and make a downie lora or whatever they are called.

>>101854443
check OP for recommendations
if a1111 works for you then you're good. (unless you want to try flux, then you'll need forge or comfy)

>>101854502
I'd wager the 5070 will also have 16gb of vram, because the 4070 ti has 16gb of vram.

>>101854518
it's also called trisomy 21
I've found it recognises synonyms better than the colloquial terms in some areas

>>101854544
Only the 4070 Ti Super has 16GB

>>101854502
I personally think the 5090 will have 28GB, the 5080 will have 20GB, the 5070 16GB, etc etc
Knowing Jewvidia having the high end all to themselves this generation, we are in for some rough times

>>101854603
There's too much money for other companies to start trying to make a dedicated chip. Whoever cracks a cheaper, high speed, high VRAM AI chip is going to be making billions.

File: file.jpg (105 KB, 1920x1080)

105 KB JPG

>>101854647
>TSMC

>>101854680
Flux can't into holograms?

>>101854686
>Flux can't into holograms?
it can if you use cfg and shit

>>101854670
zoomies wont know this one

File: ComfyUI_01346_.png (1.46 MB, 1024x1024)

1.46 MB PNG

>>101854686
had a really hard time with them, kept trying to specify and get some form of better result and it'd just make the image worse.

Here's one that coincidentally ended up looking like ana de armas though

>>101854695
cfg is on 5, guidance 8. i've seen the miku hologram example, but this one seems a bit harder to pull off

>>101854720
>cfg is on 5, guidance 8
cfg 6 + guidanceNeg 10 is when you start to see the kino

File: 1298382595.png (943 KB, 768x1344)

943 KB PNG

the perfect woman

>>101854741
two left feet

>>101854741
4 footed footjob, imagine

>>101854763
I can easily imagine that, because my dick is very large and thefore could fit all 4 feet, of course

File: ComfyUI_01358_.png (1.62 MB, 1024x1024)

1.62 MB PNG

>>101854736
i've done a bit of experimentation around the numbers (just anecdotally) and i've found that in a lot of my gens hitting cfg 6/neg guidance 10 introduces just a little bit too much instability where you have to go through more gens overall

pic rel

File: Untitled.jpg (28 KB, 512x768)

28 KB JPG

>>101854555
Thanks, but even with that I only get vaguely downy faces. Even this was cherry picked.

>>101854793
that's why you add AdaptiveGuidance on top of it, so that the cfg = 1 on the last steps cleans the picture, and it makes the gen faster

>>101854835
yes lad adaptive guidance is on there as well, i've studied all the workflows posted here by the miku professor

>>101854832
>you can do anything!
>anything?

>>101854846
>i've studied all the workflows posted here by the miku professor
you're talking with the miku professor :^)
Next lesson will be ModelSamplerTonemapNoiseTest, but I need more XY plot to find the same kind of kino as DynamicThreshold, I'm gonna sleep so I let those images rendering, see ya next day kek

>>101854544
Why would it when paypigs like you will buy the 12GB version and then shell out more for the super nigger TI with 2 more GB after a year or two

>>101854855
It all comes back around to Emma Watson 1girls in the end

>>101854868
well i gotta say cheers for all your experimentation, my gens have been a lot better overall because of them.

>>101854741
>Slaps will turn into kicks
Nope

>>101854907
thanks anon o/

So does t5 negative prompting change the way we negative prompt?

i.e I understood it as 'put anything you don't want to happen in there'. So you'd put "blurry, grainy, pixelated".

But because of the boomer prompting of T5, should I be putting "Make the image blurry, grainy and pixelated." or "Avoid anything blurry, grainy or pixelated."?

>>101854909
lol

File: ComfyUI_01364_.png (1.54 MB, 1024x1024)

1.54 MB PNG

>>101854951
Okay, quick follow up. Not an extensive test mind you but it seems to track:

https://imgsli.com/Mjg2OTY2

Looks like in the negative T5 you should be prompting that you don't want those things, rather than prompting for them (which is what we did before in the negatives).

>>101854333
This is beautiful. Prompt please.

>>101854996
>Looks like in the negative T5 you should be prompting that you don't want those things
this makes 0 sense

>>101854689
prompt?

File: ComfyUI_00205_.png (1.75 MB, 1024x1440)

1.75 MB PNG

https://civitai.com/models/641275/elle-fanning-flux-lora
New Elle Fanning lora on Civitai. I don't have much interest in celebrity 1girls but I am impressed by the results Flux can achieve with Loras trained on a relatively small data set, this one was 40 images supposedly

File: ComfyUI_01368_.png (1.46 MB, 1024x1024)

1.46 MB PNG

>>101855048
I'm not sure if you're saying that I don't make sense - or the prompting style doesn't make sense, but it's coinciding with good results.

https://imgsli.com/Mjg2OTY4

This is my negative T5 currently since Flux loves boomer prompting:

>Avoid anything blurry, low-resolution, or grainy. The image shouldn’t be pixelated, out of focus, or distorted in any way. Make sure there are no issues with stretching, warping, or watermarks, and avoid any unwanted text. Lighting should be natural, without harsh shadows, overexposure, or underexposure. Colors should be balanced—not overly saturated or unnatural. Anatomy should be accurate, with no deformed hands, missing fingers, or extra limbs. The body proportions and poses should be realistic, without any awkward or unnatural positions. Avoid noise, artifacts, or flat, cartoonish textures. The image should have plenty of detail and avoid low-detail areas or bad perspective. Backgrounds should be clean and composed well, with no unwanted objects or messy elements. Cropping should be careful, with no edges cut off, and the image shouldn’t be over-sharpened or over-smoothed. Avoid color banding, unnatural lighting effects, double images, or excessive blur. Lens flares, chromatic aberration, and moiré patterns should be avoided. Contrast should be balanced, and there should be no random objects, people, or unwanted reflections. The composition should be balanced, with no clipping issues, jagged edges, or inconsistent style. Hair should look realistic, and there should be no floating objects or unwanted borders. The overall image should have proper lighting, not too dark or too bright. Additionally, watch out for deformed anatomy, distorted limbs, misaligned joints, unnatural body shapes, awkward positioning, inconsistent muscle definition, and incorrect bone structure. Make sure body parts are connected properly and avoid unrealistic contortions or body angles.

>>101855130
>elle fanning AGAIN
what is it with people and their abundant shit taste?

Dakota fanning was always hotter.

>>101855130
>hag Elle version
stupid lora maker

>>101855120

Brunette woman with pixie undercut, fingerless leather gloves and long red acrylic spiked nails.

exactly how it looks

Is flux on comfy only for vram chads or can it offload partially to system RAM like forge? I'm trying forge but somehow the details I'm getting on my gens seem middling compared to the average images I see posted.

>>101855130
>impressed by the results Flux can achieve with Loras trained on a relatively small data set, this one was 40 images supposedly
that's how XL loras were trained as well tho

File: ComfyUI_01369_.png (1.43 MB, 1024x1024)

1.43 MB PNG

>>101854996
>>101855131

wait and wtf, positing prompting with tags in the clip also made for a better negative prompt:

https://imgsli.com/Mjg2OTcw

Why is prompting positive attributes in the negative clip & t5 giving better results? That's not how this should work.

>>101855224
imgsli description is wrong, those words are in the negative AND positive clip encoder for the right image.

File: Gen_1723487903_0002.jpg (1.65 MB, 4608x3584)

1.65 MB JPG

>>101855208
It can offload to system RAM yeah. You ideally want the UNet on GPU and CLIP on system RAM. The UNet on RAM would be extremely slow.

>>101855292
Is this pony?

>>101855215
SDXL loras trained on such a little datasets are not nearly as flexible and everything outside of closeups need adetailer

>>101855300
Yes

>>101855321
XL loras don't get more flexible with more images anyway

>>101855337
which pony mdoel specifically? the pose looks really good

>>101855355
Custom mix (2DN + RevEngine). Tho I achieved the pose using ControlNet (depth + openpose) so you can basically use any pony model.

>>101855398
>2DN
Thnx anon, I will give that model a try

File: fp071.jpg (289 KB, 1024x1024)

289 KB JPG

it's too hot

>>101855130
What are the requirements for training Flux loras like?

>>101851493
>a bitch to unfuck
already unfucked

I'm doing a Lora test on Flux and it seems quite good at learning compared to SDXL even with a limited rank 16 one. Seems to be making good progress on nudity. Also it seems to work fine with larger, diverse datasets. Might be able to train all your characters all at once along with some normalization images.

>>101855491
how difficult is it to train compared to sdxl? i've heard it's hell on earth to get sdxl refined with nsfw content.

Are loras working on NF4 yet?

>>101855522
It's still working but the learning appears to be going in the right direction and only 2500 steps in. It's so promising I'm putting together a more comprehensive test to see what happens when you mix some celebrities in with everything else. Basically I'm doing some wikiart, some random stuff from search engines, some nsfw, some booru.

>>101855574
if you're doing danbooru images, please please please make sure to fix the tags to be actual flux prompting or the whole thing is going to be ruined.

So what is everyone using? full fp16? fp8? nf4?

>>101855748
No, I don't think I will

>>101855748
>the whole thing is going to be ruined
source?

>>101855764
niggafour4

>>101855772
you will be cursed for 1000 years of bad gens.

>>101855784
its just how the dataset is. it's like if you were to train a booru style xl or 1.5 model using flux type prompting. it just won't turn out as good.

>>101855818
i thought flux supported both flux prompting and booru tag-like prompting.

>>101855784
It's all T5, so the best way to train is a variety of prompt styles. I've already seen the T5 is smart enough that it can be mix prompted between booru tags and boomerisms.

>>101855818
I gen what makes me happy. I'll look forward to your superior lora.

>>101855772
fix your fucking tags and make sure shit works correctly you gargantuan niggerbrained monkey

god people shit on vramlets but the people making all the retarded decisions are the ones with compute to train.
boy this shit gets me madder than it maybe should.

I'm debo

>>101855843
maybe the people with compute have brains
except for Pony of course

>>101855831
i mean any pony model will also support flux style prompting but it's shit compared booru style prompting.
it's just how the model was trained so you should stick to that until flux eventually gets fine-tuned to be on a booru dataset.
flux prompting also allows for a lot better prompt adherence and more detailed prompts.
try booru style prompting in flux compared to how it's supposed to be prompted; it's just worse.

>>101854443
Forg

>>101855864
no I'm debo

>>101855843
tag based prompts will work just fine with T5

>>101855784
NTA but I've seen some anime style loras on civit that looked good but then said it only works with booru prompting, which seemed to defeat the purpose of using flux. Training loras with booru tags might fuck the prompt understanding.

>think that I should use FP8 since FP16 would trigger loading between RAM and VRAM
>use it that way forever
>today think maybe I should try FP16 just to see what it's like
>first gen looks great, less fuckups compared to FP8
>AND it's not actually much slower, just like 1-4 seconds more per gen
Damn I should've tried this earlier Though I do have an extra 3060, which I offload the text encoders to, so that might be making the difference here.

>>101855940
seen enough tranime sthu

>>101855940
as much as I like typing stupid sentences, I prefer mixed prompts, and imagine coming to any conclusions with that GTX 1060 monkeys do on Civitai, most of what's over there is complete trash

File: ComfyUI_01393_.png (1.64 MB, 896x1152)

1.64 MB PNG

>>101855843
i genned your post

>>101854443
use reforge. better than a1111 in pretty much everyway.
if you want to learn comfyUI, that's technically* the best but there's a bit of a learning curve.

>>101855991
he's a real fella

File: 1697752967664118.png (1.55 MB, 832x1248)

1.55 MB PNG

alhumdullilah

File: 1714453112689275.png (1.38 MB, 832x1248)

1.38 MB PNG

discordians stay winning

>>101856061
I'm curious what prompt you used to get that backshot, because I can never seem to get my generations to take that pose. Even when I explicitly write things like "the woman has her back turned to the viewer, and is looking away from the viewer", the character still turns her head, and I can rarely ever get a backshot of her head.

File: ComfyUI_00150_.png (1.26 MB, 1024x1024)

1.26 MB PNG

Hello, /g/friends, noob here, is there any recommended video tutorial on how to use ComfyUI?

I'm currently watching one by John Malkovich, but he doesn't really explain anything, just dropped in a workflow and 6 million nodes showed up.

I want to understand.

>>101856127
NTA but "Seen from behind" works for me

>>101856175
Thanks, I'll try it

is it difficult to get gymnast/yoga poses in flux?

>>101856127

Cinematic film still of a large hologram elleff she has blue hair in a bob haircut watching over a cyberpunk city

Just a simple prompt experimenting with the Elle Fanning lora seeing how flexible it is, I didn't specifically prompt for that pose and that gen isn't really what I was looking for I just thought it looked cool

>>101856201
No

File: Screenshot 2024-08-12 154022.png (10 KB, 862x62)

10 KB PNG

>Don't have 24 gb of vram
Is it nyover for me fluxbros...

File: 51470957.png (1.58 MB, 1024x1024)

1.58 MB PNG

Good afternoon

>>101856230
Welcome to nf4 or fp8.

>>101856241
Is it worse? I know you could quantize SD models two or three times before they broken/lost quality.

File: 2024-08-12_194051_seed374(...).png (563 KB, 1280x864)

563 KB PNG

>when you play around with img2img from a completely black image and this is the first thing you get
Jesus.

>>101856267
It is worse, but not by too much. I have found that nf4 looks good to me.

>>101856282
You will be fine. Flux may do scary things but it doesn't like violence.

File: 1700524200886517.png (1.45 MB, 1024x1024)

1.45 MB PNG

>>101856230
i was on fp8 till i found out about nf4. 20 steps 1cfg, fp8 took me 150sec on a 3060 12gb, nf4 is 80sec. its not bad, but i dont think its on par with fp8.

>>101856127
"As seen standing before [object]" ... Or "As though how [character] should be seen should [pronoun] be standing before [object] with [pronoun] back/backside visible." It's something like that. It's one of those LLMisms that will work literally every time.

File: file.png (1.77 MB, 1216x832)

1.77 MB PNG

>>101855764
nf4 model + fp16 t5xxl, feels good

File: 2024-08-12_195638_seed378(...).png (782 KB, 1536x864)

782 KB PNG

You really can get some unique shots using this technique.

>>101856350
Lmao.

>>101856552
dark souls 1 abyss

Can Flux do chubby tummies without turning the girls' faces hideous?

>>101851925
You can do all these with Hunyuan. You can also do them better.

when is ponyflux?

any tips on prompting text in sdxl? im using automatic1111. i randomly generate an image and the girl's shirt says "BONGA" on it in really nice letters, i want to make it say something else

>>101856709
SDXL can't do text

>>101856383
Noted, thanks!

>>101856709
Any text you get with SDXL is just a coincidence, it has no understand of it.

File: ComfyUI_Flux_7509.jpg (312 KB, 768x1344)

312 KB JPG

>>101856709
controlnet

>>101856709
you don't prompt text in sdxl. you can however get decent text via putting text over an image and then inpainting at low denoise.
flux does text marvelously but is missing a lot of tools.

File: ComfyUI_Flux_7.png (829 KB, 832x1216)

829 KB PNG

>>101856641

>>101856532
Looks excellent, prompt please.

>>101856571
lmao was going to say this

>>101851925
SD 1.5 had soul, simple as.

>>101852526
I haven't tried embedding since 1.5 but you're probably right. I've definitely accidentally gotten some concepts in flux, it likely has the in there.

>new thing
SOULLESS
>old thing
SOVLFVL

where's the schizo who says navels are censored when you need 'em

>>101856641
>ponyflux
astracuck is irrelevant. long live the next lighting in a bottle pron tune

>>101852609
Do you have an example of this or was it revealed to you in a dream? Because I haven't seen anyone train an OC lora yet.

>>101857028
just train on gibberish search alt tags to get your random back

https://civitai.com/models/640247/mjanimefluxlora?modelVersionId=716064

anyone got a workflow to make images using that LoRa where you use multiple GPUs? I tried adding the "Force/Set Clip Device" and "Force/Set VAE Device" but keep getting OOM errors

>>101853092
Because they grew up with her.

File: _adb189a1-fbd3-451c-b850-(...).jpg (157 KB, 1024x1024)

157 KB JPG

>>101857057
I think we can probably narrow down soul vs. soulless as variety vs. homogeneity. Both language models and image models, no matter how technically proficient, become more "soulless" to users once they become REALLY good at doing exactly one thing and stop being able to do anything else. It's why people still use Dall-E despite it being a giganerfed piece of shit. You're not getting anything "out there" stylistically like this unless you specifically go out of your way to figure out whatever the fuck "this" is and download a LORA for it.

why do some loras work and some dont

How would I make a random selection with bias in ComfyUI?
ex: 75% chance of outputting "Apple", 25% chance of outputting "Pear"

>>101857071
I didn't say they were censored in and of themselves. But that area has been fucked with. You get a navel, you'll get a large belt buckle, high waisted jeans, or underwear with a large bow in front of it. This model has clearly been censored for nudity. How is that even controversial?

>>101857238
some people know how to train and some don't

>>101857247
You can't do it in Comfy lmao, use A1111

>>101856974
Looks Amazonian. Care to share the prompt?

>>101857238
Overbaking, underbaking (rare). People will just idiotically go along with default settings. Improperly curated and/or tagged datasets with dogshit images or an unclearly tagged concept/character in them.

>>101853967
2.5b is all you need bro

Ready to go with a fresh loaf of...
>>101857264
>>101857264
>>101857264

>>101857272
yes you can nigger you need a custom node for it.

>>101857248
It's not censored for nudity, it's just been very strictly filtered against nudity. SD3 did a post process lobotomy which is why it does mangled nude bodies, Flux did a preprocess filter of nsfw images which is why it simply doesn't know how to do crotches because they're not in the dataset outside of barbie dolls.

>>101857272
Retard

File: 2024-08-12_203843_seed417(...).png (889 KB, 1280x720)

889 KB PNG

And now with a blank white image. I guess this is more useful if you're going for a sketch type of gen, though it also has a unique feel to it.

>>101856571
Yeah, pretty happy it turned out that way.

>>101857247
There's no internal feature to do that as far as I know. You'll have to set up your own workflow using Math expression evaluator, ComfyUI-Logic, and some sort of string concatenate node to append it to your prompt.

Unless there's a custom node that does it for you.

>>101857275
just prompted a strong barbarian woman.

>>101857296
>>101857348
Huh, weird that no one has done it yet.
I guess I'll have to do it from scratch...

>>101857307
A distinction without much of a difference.

>>101857488
only if you're low IQ
notice how one model does mangled bodies and the other simply can't draw specific parts?