File: delux_flebo_00076_.png (1.48 MB, 1216x832)

1.48 MB PNG

>mfw

Gundam thread

Any list of known artist in flux?

File: Capture.jpg (23 KB, 667x384)

23 KB JPG

For those tinkering with ModelSamplingFlux, did you also notice that changing the base_shift value doesn't change anything?

>>101826666
well theres ummm... uh.... well if you put in some random old painter names it makes the images brown i guess..

File: ComfyUI_temp_znjeq_00005_.png (1014 KB, 1024x1024)

1014 KB PNG

Guys, can we please have a found footage thread?

File: Capture.jpg (388 KB, 2855x1735)

388 KB JPG

How can I add prompts on the XY plot?

>>101826601
Fluff all the tails!

>>101826666
I think you're vastly misunderstanding how Flux was trained and how the dataset was tagged.

>>101826767
nta, can I have a qrd or a link?

>>101826767
Any documentation on that?

>>101826776
No lol. There's a blog post.

>>101826784
This is why SAI died and the stupid faggots didn't learn shit

>>101826772
No
>>101826776
No

It came to me in a dream. The entity told me they used an LLM to tag the images. It then told me that the LLM did not know every artist and their style.

>>101826470

Flux seems rather censored. It can't do nipples right and crotches don't turn out well.

>>101826772
>>101826776
Working on it. FLOOX

okay so i downloaded ControlNet.

Now what.

>>101826802
Now control the net.

File: ComfyUI_00304_.png (1.25 MB, 1168x880)

1.25 MB PNG

You ever just

Why are these SAI goons so fucking tone deaf?

pixart used to be here

>>101826871
Don't you think that now both /sdg/ and /Idg/ use FLUX shouldn't both threads just merge?

>>101827003
They would have to come here then the schizo who hates avatar fags will have a melty.

>>101827003
bad idea, the schizos are on sdg, that's cool we can let them schizoing into a mental institution, it's like tumbler it was used to be the internet mental institution

>>101826505
TY migu

File: ComfyUI_09193_.jpg (1.16 MB, 2048x2048)

1.16 MB JPG

>>101827025
SDG gives me more (you)'s though

>>101827031
Why does the pumpkin look horny?

File: Flux_00467_.png (1.52 MB, 1024x1024)

1.52 MB PNG

>>101827003
what we really need is an /AI/ board.
then every Schizo can have his own containment thread and we can have comfy threads about any topic revolving around diffusion and AI models, prompt engineering, prompt sharing etc.

File: Flux-20240811_055350-gen-(...).png (1.2 MB, 1536x640)

1.2 MB PNG

>>101827025
tumblr wasnt bad before censorship bs

>>101827040
it's an orange

>>101827055
for me it was a good place to let the blue haired non binary crazies scream in the void, but now that tumblr killed itself the crazies spreaded all across the internet

File: CatJak_00430.png (2.37 MB, 832x1280)

2.37 MB PNG

>>101827073
true!

>>101827064
Oh that explains why it's horny. Thanks

File: SKIBIDI.png (359 KB, 423x581)

359 KB PNG

local diffusion like tea brewing!??!! LOCAL TEAFUSION DIFFUSER YEASSS YES LIKE THE DIFFUSER ON MY FUCKING TOYOTA 86 PRO COMP SPORT EDITION!?!?!?!?

I ONLY USE IT FOR GROCERIES THEN i GO Home and play on the gcomputger and go to local tea diffusion genneral ... sorry for the cuonfusion ... the diffusion COnfusion!!!!

File: Flux-20240811_050543-gen-(...).png (1.29 MB, 1216x832)

1.29 MB PNG

>>101827073
I wouldn't know about that, was just always looking for good pictures and some tumblrs delivered.

>>101827100
I mean, tumblr was good at that aswell, that's also why I still lurk on reddit, you get the news and good technolody documentation on SD in there, but I also don't forget that this is a ultra leftist place, just be careful where to put your steps

>>101827086
any new pics of your thick ass mom I should check out?

>>101827115
The internet definitely got worse after Tumblr banned porn

File: Flux-20240811_051746-gen-(...).png (2.29 MB, 1152x896)

2.29 MB PNG

>>101827086
that's gotta be one hell of an assault rifle man, or maybe he is human sized?
>>101827100
always

>>101826794
This is why proompting like an LLM gives such good results.

>>101826791
Right? Well, they don't want people getting a ton of mileage of t5 alone because then no one would pay for fine-tuning.

File: 202412202412_42.jpg (207 KB, 1296x1480)

207 KB JPG

>>101826726

>>101827149
What does this mean?

>>101827100
Very nice but can you add "(tight:1.1), ultra tight, super tight, very tight" to the prompt please?

>>101827177
Tumblr announced they would be banning all porn from the website, which caused a mass exodus of Tumblr users to the rest of the internet. Most of them took over reddit, some came here, all of them made the respective sites worse.

>>101827183
Ah I suppose that makes sense. I dont watch porn so I'm not familiar with the "community" or whatever

>>101827183
twitter was also highly infected by that, it was a way nicer place before 2018

File: Flux-20240811_061001-gen-(...).png (1.32 MB, 1536x640)

1.32 MB PNG

>>101827182
I moved on to triangular spaceships which, for some reason, are not triangular. what should "tight" do ? also, I (You) ed myself, oops

>>101827199
Tumblr used to be were e-thots would whore themselves out, but they would do it for free. Used to be a great place for amateur porn

File: 00091-60_20240211142904.png (3.9 MB, 1664x2048)

3.9 MB PNG

They really are trying to make up drama in the other thread because nobody is really posting. Just step your prompt game up dudes

>>101827236
pretty sure no one here cares other than you. You're bizarrely emotionally invested

>>101826750
I think you have it right. I've had issues when it's trying to interact with a group node and when i had two of them and one was disconnected. it's a bit of a pain but it does work since i'm running it right now.

>>101827245
Thank you for the bump

File: Flux-HR_00078_.jpg (382 KB, 1280x1280)

382 KB JPG

>>101827216
>what should "tight" do ?
The world may never know

>>101827250
you managed to make the prompt work on the XYZ plot node?

>>101827100
>>101827255
cool cars

Trying to test out that new canny controlnet for flux, but I keep getting keys errors even though I switch the git to git xlabs_flux_controlnet. Is there something else I need to do? Updating just switches it back to main.

>>101827253
I couldn't care less which thread is bumped more, because unlike you I don't have an emotional investment. Fucking pathetic, lmao

Blessed thread of frenship

>>101827288
Stop crying

>>101826726
Final boss of Space Invaders for N64
https://youtu.be/TT5sevJVX5w?si=yyd4zfF2bonh4n_B&t=301

zzzzz... mimimimi.... zzzzz.... mimimimi...

File: FLUX_00101_.png (3.86 MB, 2016x1152)

3.86 MB PNG

>>101826726

>>101827294
cognitive biases thicker than a Louisiana woman's ass

ChuckMcSneed uploaded full FLUX dev model so it's available without needing to make a HuggingFace account.

https://huggingface.co/ChuckMcSneed/FLUX.1-dev/tree/main

>>101827317
How big is a louisiana womans ass normally? Is it a good kind of big or a morbidly obese kind of big?

>>101827332
i have never wanted to download something less

>>101827303
This image sucks
>Plastic skin
>Nonsensical clothes

It's like the first time you ever tried typing naked woman into flux.

>>101827336
He's a crying west coast schizo that doesn't know anything about the outside world

>>101827332
I can only imagine how slow the full model weights are

>>101827336
like two big black beach balls tied together bouncing around

https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981
Game over for ComfyUi

File: Flux-20240811_062947-gen-(...).png (1.74 MB, 896x1152)

1.74 MB PNG

>>101827350
dangerous.
>>101827341
cut the man some slack! maybe it is his first horndog gen.

>>101827357
>(i) NF4 is significantly faster than FP8. For GPUs with 6GB/8GB VRAM, the speed-up is about 3.5x to 4x. I test 3070 laptop (8GB VRAM) just now, the FP8 is 8.3 seconds per iteration; NF4 is 2.15 seconds per iteration (in my case, 3.86x faster)
>(ii) NF4 weights are about half size of FP8.
>(iii) NF4 may outperform FP8 (e4m3fn/e5m2) in numerical precision, and it does outperform e4m3fn/e5m2 in many (in fact, most) cases/benchmarks.
What the fuck??? how does a 4bit quant perform better than fp8??

>>101827369
this is a very good artstyle

>>101827345
I'm not that anon nor am I American so I simply have no exposure to a louisiana womans ass
>>101827351
I don't like this image

>>101827332
Based.

>>101827357
holy, i might actually use flux before the upgrade

File: ComfyUI_03108_.png (1.5 MB, 896x1152)

1.5 MB PNG

>>101827383
>I don't like this image
he certainly does!

File: Flux-20240811_062906-gen-(...).png (1.55 MB, 896x1152)

1.55 MB PNG

>>101827357
please. illyas gonna burn himself out real good and abandon ship before you know it.
>>101827377
ty, well thank lotte reininger

>>101827338
It's very useful if you want to make your own quants or need the exact model config

File: 1699098636467287.png (392 KB, 579x540)

392 KB PNG

>>101827260
i thought it was fine since i did samplers. i got the same error when i tried to use it to edit the primitive prompt node. i tried a string function node and that worked though. thats a pythongosss addon though. im too lazy too figure out if there's a base string node that would work. it can be finagled i guess though.

>>101827419
>illyas gonna burn himself out real good and abandon ship before you know it.
look at this shit, if this is true it's a fucking revolution >>101827375
4 times the speed and better precision than fp8? NO FUCKING WAY

>>101827341
You're retarded.

>>101827369
So are you.

>>101827424
you couldnt make a donald trump dog though

File: vramlets are saved.jpg (108 KB, 858x811)

108 KB JPG

>>101827357
there is even more optimizations further in
>2 min flux dev 20 steps 3050

>>101827357
>>101827428
Comfy can't run the bnb-nf4?

>>101827332
I kneel

>>101827459
I don't think so, if he doesn't implement this he's dead, everyone will jump to the forge ship

>>101827459
Dunno. Doesn't take long to add though.

>>101827375
I'm >>101815411
And that's what I used. Very easy to implement in comfy or whatever.

File: Flux-20240811_062155-gen-(...).png (470 KB, 896x1152)

470 KB PNG

>>101827428
gonna wipe my old forge install and test it out tomorrow, sounds very very exiting desu. but its 6am, not now

>>101827375
Excuse me? Is this magic or something? Everyone will be able to run forge now and at high speed

>>101827369
>>101827491
Into the collage

>>101827486
>And that's what I used. Very easy to implement in comfy or whatever.
you used nf4? does the quality really equivalent to fp8?

File: ComfyUI_03114_.png (3.9 MB, 1792x2304)

3.9 MB PNG

I call this one, "Mama"

Alright bros, can someone share a catbox for a flux workflow that has all of the new shit? Loras/etc/etc.. thanks bro.

>>101827513
would, even with that third leg. gonna get in the way sooner or later tho
>>101827450
lol

>>101827375
>2 times smaller than fp8
>4 times faster
>better precision than fp8
what? he's trolling or something?

>>101827511
Well, the 4bit quants are better than torch's fp8 dtypes. I tried every single quant option I could find.
It wasn't better than bf16 in my tests though.

>>101827556
>Well, the 4bit quants are better than torch's fp8 dtypes.
like for real the image quality of nf4 is better than fp8_e4? this is INSANE

This is what i love about this community, all we needed was a big fuck you model for everyone to turn their attention towards optimization. Can you imagine how delayed these optimizations would be if we relied on stability's slop?

File: ComfyUI_temp_ouopa_00019_.png (111 KB, 960x1088)

111 KB PNG

>>101827549

>>101827357
I just threw my comfy ui in the bin. It's over.

>>101827357
https://huggingface.co/lllyasviel/flux1-dev-bnb-nf4
>T5xxl in fp8e4m3fn
>not fp16
NOOOOOO NOT LIKE THIS, FUCK OFF

File: Flux-20240811_050838-gen-(...).png (971 KB, 1152x896)

971 KB PNG

>>101827556
thats to be expected, yes. but damn good times ahead mofos!
>>101827568
yes. see krauts delivered

Didn't lllyasviel make it so you can't use his code for profit?

>>101827581
>thats to be expected, yes. but damn good times ahead mofos!
this is insanity, you halve the size, get a 4x speed increase and you even get a better quality? holy fuck this is magic

>>101827580
i am sure we will get some option to load the T5 separately in fp16 at some point.

>>101827587
Chang don't care, what you gonna do 'bout it?

>>101827599
I don't know why he doesn't separate the three of them, one download for flux, one download for the clip encoder and one download for the VAE, merging everything isn't convenient at all, what if I want to change the VAE? well I fucking can't with this method

>>101827606
Very nice

>>101827375
does someone have a live lykon reaction? their last cope was that their sd3 2.5b model was made for vramlets

File: dogolie1.jpg (281 KB, 1024x1024)

281 KB JPG

>>101827357
>>101827375
We are SO back

>>101827625
>their last cope was that their sd3 2.5b model was made for vramlets
you're kidding, right?

>>101827375
>I test 3070 laptop (8GB VRAM) just now, the FP8 is 8.3 seconds per iteration; NF4 is 2.15 seconds per iteration (in my case, 3.86x faster)
that's not a good comparaison, the 3070 doesn't have enough VRAM for fp8 so its naturally slow because of the ram offload, a better comparaison would be a 16gb or a 24gb 3xxx card

>If your device is GPU with GTX 10XX/20XX then your device may not support NF4, please download the flux1-dev-fp8.safetensors.
Fug.
Download cancelled

>>101827332
Kek
The only discussion is a desu image by desuanon and Chuck pinned it
https://huggingface.co/ChuckMcSneed/FLUX.1-dev/discussions/1

File: ComfyUI_03125_.jpg (878 KB, 1792x2304)

878 KB JPG

mama and sanic? oh lawd have mercy

File: media_GUPmuEVXwAA2kb8.jpg (106 KB, 1024x1024)

106 KB JPG

>>101827635
no he's 100% right lmao
https://xcancel.com/Konan92_AI/status/1820552363482206558#m

File: download.jpg (17 KB, 299x168)

17 KB JPG

>>101827641
>If your device is GPU with GTX 10XX/20XX then your device may not support NF4, please download the flux1-dev-fp8.safetensors.
the 3xxx series sales after that announcement

>>101827641
always read the fine print first. good man.
BOY AM I GLAD I GOT A 30XX

>>101827647
/OurGuy/

File: download (1).jpg (10 KB, 279x181)

10 KB JPG

>>101827641
even though I have a 3090 card, I want that 4x speed, and better precision than fp8_e4, you goddam right I'm downloading that model

File: ComfyUI_03127_.png (1.39 MB, 1152x896)

1.39 MB PNG

no mama!

remember people scream and crying about how flux wouldn't get any faster or better, was untrainable/untunable, and would never replace existing models

lol

>>101827375
Comfy you better implement this quickly, or else you can say goodbye to your career

>>101827711
I wont stop until hyper realistic porn tailored to my unique and over the top fetishes is beamed into my eyeballs at 166 fps on a 1080

he aint even here bro

>>101827719
You forgot to mention in stereo 3D

File: haha-im-in-danger-meme-2.png (148 KB, 367x374)

148 KB PNG

>>101827375
>be me Comfy
>make spaggheti shit Ui
>people are forced to use it because forge is having a burnout again
>feelsgoodman, here goes my monopoly!
>oh wait...
>forge is back
>so back
>so fucking back
>implement nf4
>4x speed improvement over fp8
>2 times lighter than fp8
>better precision than fp8
>uh oh...

>>101827467
He can’t copy the code tho, it’s clearly stated

File: 1721340864336626.gif (499 KB, 492x318)

499 KB GIF

>Think it's game over with my puny vramlet 10GB 3080 that's been barely hanging on with the current models and I'm stuck with SDXL until 5090 comes out next year.
>Mfw autistic techno wizardry allows me to play with the kickass new model without problems

>>101827741
lmao, forge made the perfect move there

>>101827717
Link it in the tranny thread so he actually sees it.

>>101827745
same thing here but with a 3070, what a fucking cursed generation of cards

>>101827599
You can just connect the model nodes, and use t5&clip and vae from separated nodes.

>>101827756
Debo does that already

>>101827761
i am talking about forge

>>101827741
Using a simple quant of a model for your own software isn't something that would be protected like that. Even if it would, it would need to be changed depending on your software

>>101827741
he can't copy the code. but we can. aha
now could you please stop polishing illiays dick?
>>101827757
rtx3070 got only 8gb of vram, right? sucks because a 12gb card can fit the text encoder in a dual card setup

>>101827741
>He can’t copy the code tho, it’s clearly stated
he did copy Comfy's code though, he's a bit hypocritical there

>>101827775
># Single File Implementation of Flux with aggressive optimizations, Copyright Forge 2024
># If used outside Forge, only non-commercial use is allowed.
>Comfy is commercial

>>101827375
So, it's on the master right? we can just update the repo and that's it?

>>101827375
I thought flux was the best shit we ever had in the imagegen ecosystem, turns out this nf4 shit is a close second, this man is simply brillant

>>101827357
i get 30 sec down from 60 sec on comfy on 3080. but it is 4bit so that's to be expected. it comes down to how much quality was lost.

>>101827833
>it comes down to how much quality was lost.
he said that the quality is better than fp8 >>101827375
>>(iii) NF4 may outperform FP8 (e4m3fn/e5m2) in numerical precision, and it does outperform e4m3fn/e5m2 in many (in fact, most) cases/benchmarks.
what do you think?

>>101827814
>Using a simple quant of a model for your own software isn't something that would be protected like that.

also anon had done it before so shouldn't be too hard >>101827486

>>101827833
compare the 2 under the same settings on forge, the uis don't have the same interpretation

>>101827357
never used forge, but I might because of that interesting nf4 thing, you just use the one click package and it will work right away like on comfy?

>>101827870
https://github.com/lllyasviel/stable-diffusion-webui-forge?tab=readme-ov-file#installing-forge

>>101827833
can you imagine adding tensorRT on top of that? 15 seconds flux dev gens here we go

>>101827031
skill issue

>>101827887
>dalle tier output gens in 15 seconds
I thought I would never see something like that in my lifetime

>>101827898
and that is on a 3080, a 4090 is probably going to spit out a gen every other second lol

Custom nf4 support for Comfy will be out later today.

>>101827865
Is not about the quant model but the gen results, obviously webui generates better quality outputs than comfy
Obviously comfy will implement it somehow but the quality of the gens/performance will be different if now worse than forge

File: images.jpg (5 KB, 225x225)

5 KB JPG

>>101827912
>later today.
I think I know what keyboard touches Comfy will be using the most today

You guys using dev or schnell?

>Imagegen community finally being forced to work their ass off to optimize and compress as much as possible like local LLMs have been doing all year
beautiful

>>101827938
schnell is deprecated because of nf4 now, flux will have great speed even at high steps

>>101827929
>obviously webui generates better quality outputs than comfy
how do you mean?
I use neither.

File: Flux-20240811_050715-gen-(...).png (1.08 MB, 1216x832)

1.08 MB PNG

>>101827912
noice. no rest for the wicked.
>>101827938
dev

>>101827942
I'm a LLM fag and I'm having a hard time to believe that nf4 is better than fp8, no one use nf4 on the language models

>>101827357
Never used forge before. It looks like A1111 except good. Is that a reasonable assessment?

>>101827956
cool e9

>>101827961
e5 was supposed to be better than e4 yet on imggen it's the opposite

>>101827966
it's basically A1111 but he made it better and faster because A1111 is a lazy ass

>>101827961
Well I guess we will find out soon, since we can begin testing it

>>101827961
LLM quant methods already use different precision per weight group, it's not dumb qunats like it used to be/like image models were using

>>101827961
I mean, LLMs have all kinds of elaborate quantization that makes things like Q5 comparable to fp8, and image generation is a different beast, so I find it plausible... although it is a little hard to believe it's actually better, rather than about the same

DynamicThresholding won't work on forge if you add adaptiveGuidance right? because the code he changed was only for ComfyUi
https://github.com/mcmonkeyprojects/sd-dynamic-thresholding/commit/5d63447afbc44b377f706a5eb0430f85791dcf30

>>101827966
just install it and see for yourself. very similar to a1111 but well enhanced. I cant say anything about the latest version tho.
>>101827970
could gen those forever, flux so good for that stuff. try motorcycles!

>>101827961
i don't believe 4 is better than 8 either but 4 is a well known sweetspot for llms. it's 3 where it starts to fall apart and 2 where it's unusable. i wonder if imagegen will be diff.

>>101826887
Pretty colors

>>101828002
>flux so good for that stuff. try motorcycles!
dont have it installed, this is my first time looking at this thread. the car just caught my eye
could you try a yamaha FZ750? thats my bike

>>101827961
>This is the NF4 state dict, you can see that each weight is stored in 6 different tensors, and these tensors are in different precisions including float32, uint8, etc:
it's likely it's the case, it's using float32 on very important weights, unlike fp8 who puts 8bit everywhere

>>101827978
>LLM quant methods already use different precision per weight group, it's not dumb qunats like it used to be/like image models were using
does that mean nf4 lora training is possible? since it's supposedly higher precision than fp8?

>>101827031
Damn Laura Palmer grew old

WE ARE SO FUCKING BACK BROS
Also I had been holding off on moving to the new forge on Gradio 4, but this shit seems faster even for SDXL gens. All the extensions I use work to.. kino

File: ComfyUI_00411_.png (1.44 MB, 1024x1024)

1.44 MB PNG

>>101828024
NTA but

File: 00002-3072641075.png (1 MB, 1024x1024)

1 MB PNG

nf4 forge -> 28 seconds on my 3090 for 20 steps + 1024x1024, that's 2 times faster than what I have on ComfyUi

>>101827961
It's because FP8 isn't a quant, it's just a half (quarter?) precision copy of the weights

i was doubting flux majorly but the recent developments seem incredibly promising. i really want to train a lora on 24gb but i'm going to hold off for a few weeks until it becomes more standardized.

>>101828082
all we need now is the t5 on fp16 instead of the bundled one in fp8. can you load the t5 separetly?

Once comfy implements it I'm switching back desu

>>101828095
I don't think we can, which is dumb, I want the 3 (image model + text encoder + vae) separated

File: Flux-20240811_074145-gen-(...).png (981 KB, 1152x896)

981 KB PNG

>>101828024

>>101828097
the main advantage will probably be the fp16 t5 if that isn't already implemented on forge

I'll wait on comfy, his UI has the controlnet support which I was just fucking around with before this dropped. I still get key errors in forge.

>>101828077
i love you
exhaust is on the wrong side unless thats an 85 :^)
also really interesting that it put the indicators molded into the fairings like a modern bike
even more interesting is that it got the alternator cover right, which the FZ is one of the few bikes to even have, but put it in the wrong place

the preview doesn't seem to work on forge, I've looked at it it's activated, do you have the same problem?

File: Flux-20240811_074417-gen-(...).png (938 KB, 1152x896)

938 KB PNG

>>101828056
sdxl faster too, wtf?
>>101828082
oh wow
ok I am exited now

>>101828082
I get about 40secs per image on a 4080, so will see once forge installs if I save any time without losing too much quality

can we use lora in forge though?

>>101827961
language != image
mathematically different.

>>101828082
>nf4 status (nf4 flux + fp16 vae + fp8 t5xxl + fp16 clip_l)
>12.8gb when loading the nf4 model
>13.9gb when making the picture
>16.6gb VAE spike at the decoding of the picture
https://huggingface.co/lllyasviel/flux1-dev-bnb-nf4

>>101828151
why didn't he put the fp16 t5xxl instead... that's retarded

Where do I chose my checkpoint folder in Forge? I haven't used Gradio in forever but I remember you could do it in A1

File: 00005-2245008562.png (1.16 MB, 1024x1024)

1.16 MB PNG

>>101828164
fp8 t5xxl completely destroys the text, FUCK

>>101828164
He'll do it eventually, trust the plan.

This thing has never seen a vagina but it just so happens to know how to outline the basic shape of one in lace kek

https://files.catbox.moe/lm6u4v.jpg

>>101828082
wait a minute. I get low 30 seconds for 20 steps and 44 with a neg on a 3090 and comfy. (@80%power) your times are off

is it worth even trying any of this with a 3070?

>>101828184
I have activated a temp limit on my 3090, that's why it's not faster than yours

>>101828188
Yes, this update was practically made for 3070s

>>101828188
>3070
if that's the 16gb one it'll work for that one >>101828151

>>101828164
Image gen people are fucking retards is why. I would be shocked if any of them get any mileage out of t5 other than the boomer prompters.

>>101827943
Need separated nf4 only model though. Already have t5 and vae, don't want to waste space.

>>101828202
show me your gens

>>101828192
my card never goes above 66 or so even at full power, good airflow
>>101828176
help me understand your (fantastic) gens. why the double image? LRF Full SBS?

>>101828203
true, a separation is needed, I want to change VAEs and shit, having to redownload the same t5 for each new flux model would be a retarded move

File: Capture.jpg (153 KB, 1814x1362)

153 KB JPG

I hope he'll make a distilled NEGATIVE CFG scale too, because that shit is needed to make magic happen
https://reddit.com/r/StableDiffusion/comments/1enm9og/discovered_by_accident_a_trick_to_make_flux/

File: test.png (1.5 MB, 1024x1024)

1.5 MB PNG

Hello. I remember reading here a couple of days ago that there was a way to increase CLIP influence over the T5 text, while using both.

Can someone post the comfy flow that does that?

>>101828169
plz respond. Do I really have to copy the models over to the forge directory?

File: Flux-HR_00123__LRF_Full_SBS.jpg (693 KB, 2560x1280)

693 KB JPG

>>101828215
>help me understand your (fantastic) gens
Thanks, I meant to post this one though. If you open the image up, zoom out a bit (the more you zoom out the easier and more comfortable it will be to view) then cross your eyes so the two images overlap it's in stereo 3D.

>>101828226
Bring it up to him on github and he'll probably do it, he has been on a rampage lately.

>>101828230
ur a big boy yu can figure it out

File: ComfyUI_00430_.png (1.51 MB, 1024x1024)

1.51 MB PNG

>>101828108
Yeah, I think motorbikes are difficult since there'll be so few of them in the training data. You can see that more common models appear closer to the real thing.

Trying Flux but I only know how to booru prompt

do you still slap greg rutkowski on that shit

>>101828170
first extension is one to load the fp 16 t5 kek

>>101828266
go to copilot and ask for it to enhance your prompt for dalle 3, is not ideal, but it is a starter.

>>101828266
just ask chatgpt to generate a T5 image prompt for <thing you want>

>>101828170
it's hard to know if the quality is similar to comfyui fp8 dev + fp16 t5 when we're forced on using the shitty fp8 t5... good job forge, you autistic bastard :'(

>>101828151
I don't recommand using the async method, this shit makes the VRAM spike so hard, I got a OOM at some point juste because I wanted to change the resolution

File: test2.png (1.64 MB, 1024x1024)

1.64 MB PNG

>>101828229
Please, help
How do I make CLIP matter more than the T5 text, while still using T5?
I saw the other day someone testing different artstyles with Miku by changing CLIP settings, but I can't find it

>>101828325
https://desuarchive.org/g/thread/101738379/#101740791

File: async.jpg (60 KB, 899x186)

60 KB JPG

>>101828319

>>101828107
Delete everything expect your models folder and pull from scratch. Git pull updating broke at some point, though its fine since the last couple of versions.

>>101828249
I literally can't there is no setting for the checkpoint directory that I can see

>>101828198
there is no 16GB 3070

>>101828179
Well I have also never seen a vagina yet I know what they look like.

>>101828360
dude, open webui-user.bat with notepad;
set COMMANDLINE_ARGS= "c:\path\to\your\friggin\checkpoints"
there are other ones too, go find them.
>>101828331
some good images in that thread

File: 00004-3290183669.png (1.59 MB, 1152x896)

1.59 MB PNG

speed difference doesn't really seem that significant, unless you're comparing cfg 1 in forge with a cfg >1 in comfy
idk, i'll play around more in the morning

File: file.png (967 KB, 1024x1024)

967 KB PNG

~8s/it on a 2080 ti, fits in the 11GB of ram at 1024x1024

all this talk of fp8 T5 being shit and no multiple seed side by side

>>101828331
thank you
it seems you just have to lower the guidance?

>>101828398
>speed difference doesn't really seem that significant, unless you're comparing cfg 1 in forge with a cfg >1 in comfy
yeah, on cfg > 1 it takes about 50 seconds on my 3090, it's 1 mn on comfyui for the fp16 one, still it's faster but if the quality is worse then I'm not buying it, hard to do any comparaisons though we don't have the fp16 t5xxl to run

>>101828407
>all this talk of fp8 T5 being shit and no multiple seed side by side
how can we even do a comparaison? forge calculates shit differently than comfyUi, so it will have a different picture regardless on everything else, the only way to make a good comparaison would be to stay on forge and do a fp8 t5 vs fp16 t5 but at the moment we don't have the both of them in our hands

>>101828436
I got that when I used "async", never touch that shit

>>101828435
>forge calculates shit differently than comfyUi
why use forge at all for the comparison, anon
you do know you can load T5 fp16 and fp8 in ComfyUI, right?

>>101828407
you can compare fp8 t5 and fp16 t5 on comfy right now, it fucks with text and finer details.

>>101828449
no it doesn't
prove me wrong

File: file.png (847 KB, 1024x1024)

847 KB PNG

>>101828170

>>101828447
>you do know you can load T5 fp16 and fp8 in ComfyUI, right?
and? the embedings made by the text encoder will be used differently on forge vs comfy, that means that maybe t5 fp8 works fine for fp8 comfyui but t5 fp8 won't work well for nf4 forge, it's way more complicated than what you actually think

>>101828449
>you can compare fp8 t5 and fp16 t5 on comfy right now, it fucks with text and finer details.
I made 3 thousands gens on ComyUi (fp8 dev + fp16 t5xxl), not a single time I got something this fucked in the text like I actually see on forge after like 3 tries >>101828170

So you make your own conclusion, or else nf4 is just worse, or else it's the fault of fp8 on t5

File: fugr.jpg (55 KB, 962x339)

55 KB JPG

>>101828436

I get that same shit
Deleted venv and let it build it again.
Now the fucking thing doesn't even start, because pic related.
Had to take old venv out of the trash bin and put it back to get the program to even launch, but I'm left with that original error.
Also after rebuilding venv it said that Torch is not able to use GPU, so I had to put the skip cuda check in there.
Everything else generates stuff just fine except Flux.
I fucking hate updates, there's always something that fucks it up.

>>101828462
no, anon, it is not more complicated
if forge is using the embeddings "differently" THEN IT IS DOING SOMETHING VERY FUCKING WRONG

>>101828502
so you don't know that forge calculates shit differently than comfyUi? what a fucking new fag, get the fuck out of there

>>101828508
>calculates shit differently
lol, lmao even
anon, you're way out of your depth

File: 00000-3864446906.png (1.8 MB, 1344x768)

1.8 MB PNG

Flux working fine for me in Forge kek

>>101828518
go for it anon, make a gen with fp8 (dev + t5xxl) on ComfyUi and on Forge with the same exact settings, and you'll see you won't get the same picture at all

seems that we cant use lora in forge with the quant model

>>101828549
What does that mean do we need to disable the extension?

File: gigachad_fake.jpg (54 KB, 680x680)

54 KB JPG

>be me
>gtx1060 6GB
>using forge Animagine xl 3.1 for a month
>even used a minor one-click-easy-installer of them
>dunno and cant tell the difference of fp8 and 16
Soon I will be on winner side again.
Not sure now though.
Let's go, hackers.

nf4 left fp8 right
most come out very similar i just kept adding stuff till it actually differentiated a little. seed 2, euler, simple

https://files.catbox.moe/o5w1gj.png

photograph of a chinese woman standing in a doorway. her left hand is making a peace sign next to her face. her right hand is holding a sign that says "how long can I make this sign until one of these two models breaks". she has green hair and is wearing an army helmet. she has an ak-47 strapped on her back. she is wearing a white blouse and black jeans. there is a christmas tree in the background.

File: nf4 forge vs fp8 comfy 4080.jpg (1.56 MB, 3999x1999)

1.56 MB JPG

It's faster but not by a lot. The biggest thing it does is not overflow my VRAM into RAM so my computer is still useable. The outputs are very different, but still pretty good.
4080 32GB RAM. both had warm up time.

>>101828608
looks to be a big enough difference for anon to sperg out about it desu

File: LIAR.png (426 KB, 3456x963)

426 KB PNG

>>101827375
>NF4 is significantly faster than FP8. For GPUs with 6GB/8GB VRAM, the speed-up is about 3.5x to 4x. I test 3070 laptop (8GB VRAM) just now, the FP8 is 8.3 seconds per iteration; NF4 is 2.15 seconds per iteration (in my case, 3.86x faster)
This fucker lied to us, there's barely a difference between the nf4 forge speed and the fp8 ComfyUi speed (20 steps, CFG = 3, RTX 3090)

File: Screenshot 2024-08-11 183632.png (138 KB, 1272x1527)

138 KB PNG

>>101828611

File: ComfyUI_04691_.png (1.33 MB, 1024x1024)

1.33 MB PNG

>>101828608

Broken gun barrels on nf4 may be a deal breaker. Isolated, unlucky, or consistently bent gun barrels? Cherry pick better examples.

>>101828631
The majority of speedup is for 8GB VRAM I guess

>>101828608
>nf4 left fp8 right
both on forge? or the fp8 on comfyui?

>>101828631
he mentioned it on a gpu with 8gb that was loading a 12gb model. however you still get a speedup regardless

>>101828642
>Broken gun barrels on nf4 may be a deal breaker. Isolated, unlucky, or consistently bent gun barrels? Cherry pick better examples.
or else it's nf4's fault or else it's because of the fp8 t5, we don't know yet

>>101828653
if the quality is worse I don't want that speedup

>>101828631
>3070 laptop
>NF4 is 2.15 seconds per iteration

Reality:
>3090
>2.7 seconds per iteration

Wow he really stepped up his lies this time.

>>101828663
quality is better thoughbeit

>>101828648
was forge only. wanted to make sure everything was same.

>>101828608
do a fp16, so that we can see which one is the closest to the real deal

File: Screenshot 2024-08-11 185255.png (762 KB, 1335x669)

762 KB PNG

Text test. Suddenly became German.
nf4 and fp4 look identical

File: ComfyUI_00116_.png (1.07 MB, 1024x1024)

1.07 MB PNG

>start installing forge
>unskippable download of a sd1.5 model

>>101828608
Here's a imgsli to see better the differences
https://imgsli.com/Mjg2NDA5

>>101828710
That annoyed me too, and it's a garbage model too.

making collage

>>101828539
can't load the unet, encoders and vae separately in Forge so I can't test using the same files
no, I'm not downloading his all in one fp8 file

Here we go...
>>101828788
>>101828788
>>101828788

File: cat_angry.jpg (122 KB, 631x678)

122 KB JPG

sdxl+decent prompt when!!

Anyone else getting this?

\webui\backend\attention.py", line 314, in attention_xformers mask_out[:, :, :mask.shape[-1]]
 = mask RuntimeError: The expanded size of the tensor (1) must match the existing size (64) at non-singleton dimension 0.  Target sizes: [1, 256, 256].  Tensor sizes: [64, 256, 256] 
The expanded size of the tensor (1) must match the existing size (64) at non-singleton dimension 0.  Target sizes: 
[1, 256, 256].  Tensor sizes: [64, 256, 256]

This is even on a fresh install

>>101828687
i dont have fp16 with baked in clip vae which forge uses. as i suspected comfygen was totally different.

>>101828698
>fp4
say what

>>101828720
>calling RealisticVision a garbage model
You watch your mouth.

>>101828805
forge is such a snake...

he claimed insane speed improvement that doesn't really exist >>101828631

he gave us a model that has fp8 t5, so we don't know if the image fucks up are because of the quantized text encoder or nf4 >>101827580

and we have no way on knowing if nf4 is actually better because we can't test out the fp16 on his software

THATS FUCKING GREAT IF YOU ASK ME

>>101828833
holy fucking cope, may as well throw on a trip comfy faggot kek

File: file.png (13 KB, 955x126)

13 KB PNG

>>101827641
rtx2080 vramlet here, it definitely works for me, nf4 vs fp8

File: Screenshot 2024-08-11 194141.png (922 KB, 1089x1062)

922 KB PNG

https://imgsli.com/Mjg2NDEy
The difference between fp8 and np4 is minor

>>101829176
Actually from these examples np4 is better.

>>101829176
NP4 is noticeably better

File: Capture.jpg (447 KB, 3840x1749)

447 KB JPG

>>101828608
>seed 2, euler, simple
>photograph of a chinese woman standing in a doorway. her left hand is making a peace sign next to her face. her right hand is holding a sign that says "how long can I make this sign until one of these two models breaks". she has green hair and is wearing an army helmet. she has an ak-47 strapped on her back. she is wearing a white blouse and black jeans. there is a christmas tree in the background.
can't manage to get the same output as you anon