File: delux_flebo_00005_.jpg (304 KB, 768x768)

304 KB JPG

>mfw

>>101857264
thx baker man

cant get the nude loras to work but i can get other ones to work.

Blessed thread of frenship

PINGAS

File: 2024-08-12_203928_seed418(...).png (1.12 MB, 1280x720)

1.12 MB PNG

Another unique gen. Though shame about the sword again.

File: 565246505.png (1.98 MB, 1024x1024)

1.98 MB PNG

>>101857362
nice, I like how the right looks like a child's drawing and yeah sword is hit or miss sometimes

File: 2024-08-12_204359_seed424(...).png (1.13 MB, 1280x720)

1.13 MB PNG

Ok, I think I'm going to keep using this method from now on.

whenever I change my prompt with flux the gen takes significantly longer per iteration than the last one. (Note I am not talking about T5 encoding time, but about it/s). what's going on? where's the leak?

ComfyUI ofc

>>101857362
is it just me or some styles have way worse hands than others? "drawn by" gets me mangled hands often if the character is holding something. meanwhile if i leave it on the default flux slop style the hands are always perfect.

File: 2024-08-12_204702_seed428(...).png (1.59 MB, 1280x720)

1.59 MB PNG

Using non-solid color images now. I just drew some random lines in paint with dif colors lol.

File: 2024-08-12_205004_seed432(...).png (1.44 MB, 1280x720)

1.44 MB PNG

>>101857468
That's normal for most models unfortunately. It learns to associate bad hands with drawn styles, because hands are more often drawn badly.

File: 05131.png (1.13 MB, 768x1024)

1.13 MB PNG

I've done a shit load of tests and I've come to the conclusion that they didn't completely scrub artists from their data. However, they either poisoned the data to decrease the similarity OR the model is flawed due to training and blends the styles too much. You can still reference styles through artists names, it's just imperfect. So major artists were intentionally scrubbed, but less prominent artists are still in the data but imperfect.

>>101857343
That's because the model is censored. Good nude LoRAs will be a nightmare to train. I would just assume they will never really work.

File: file.png (1.8 MB, 1024x1024)

1.8 MB PNG

>>101857129
anyone?

>>101857544
It was revealed to me in a dream.
See you in 2 weeks

File: Capture.png (7 KB, 303x108)

7 KB PNG

What does this setting do and why does it speed up my gens 2x over leaving it default? Yes I'm retarded

>>101857540
I asked both chatgpt and gemini to caption an image of 2B without mentioning 2B and the result was what looked like really bad fucking fanart of 2B. Way way way worse than what was produced by SD1.5 with identical generation parameters.

>>101857544
not in my tests that took 3 hours of training

>>101857583
faster math done slightly wrong

>>101857247
>real world use cases
>comfyui

>>101857645
Is it still better than using the fp8 checkpoint or should I swap to that?

>>101857583
smaller numbers, can fit all at once in the GPU, computes faster because no waiting for moving numbers to and from RAM

File: 2024-08-12_205913_seed444(...).png (1.37 MB, 1280x720)

1.37 MB PNG

Ok so I was using high denoising which made colors blend together so ultimately those random lines were just noise. With high noise you need to use large blocks of colors to have an impact, like so.

>>101857686
If your system can handle the default use default it will always be better, if you have 16gb vram oress use e4

>>101857736
that's not what was asked, retard

>>101857664
You can do it in comfy with math nodes.

/ldg/? Is this new? Is it because of the downfall of Stable Diffusion?

>>101857751
Yes it is. The fp8 checkpoint is the full model but is using e4. If you can handle it use the default model.

>>101857760
there was a split between sdg and ldg. sdg still exists

File: ComfyUI_Flux_7543.jpg (225 KB, 768x1344)

225 KB JPG

>>101857760
a lot of new models appeared in the past months, which made stable diffusion not the only choice for local use

So flux knows where nipples are placed and something should go there (it's just wrong) unlike genitals which are barbie doll smooth. So what the fuck did they do to their training set?

>>101857802
Guys have nipples so it learned from that but it's wrong to have girl nipple

>>101857790
no, retard, "default" is one of the options in the drop down. no one calls the fp16 model "the default"

>>101857802
Based on a discussion I read on this a while ago, it's something along the lines of Flux Dev and Flux Schnell being distilled versions of Flux Pro with 'a strong hardcoded negative prompt'. I assume prompting female nipples fights against the 'baked in negative conditioning' and confuses the model.

>>101857751
>help me i'm a retard
>you didn't help me, you retard

>>101857839
Semantics. Doesn't change my answer

>>101857851
retard
>>101857867
it's not fucking semantics, please kill yourself

>>101857870
You're the retard mate
The fp8 checkpoint isn't a different model, it's the same model with fp8 weights and text encoding. It's the same thing as using the e4 weight type and find t5xxl. There's no difference.

>>101857911
Fp8 t5xxl, not find, phone posting

>>101857841
Ahhh that would fucking explain why the behavior is aggressively weird.

>>101857911
buddy you're not even wrong, that's just not the conversation at all, kindly shut up

>>101857851
I asked the original question I'm not the guy sperging out. I appreciate the answers

>>101858045
I like the ice crystal one

>>101858038
but you weren't given the right answer at all
there is no difference between using the fp8 checkpoint and loading the fp16 checkpoint with the fp8_e4m3fn weight_dtype so no point downloading the fp8 one

>>101858086
m8 how about you deepthroat a dildo and shut the fuck up eh? I had a bad day and now I gotta come back to your bitch ass yapping.

>>101858134
shut up retard, you can't even correctly parse what people are asking you stupid fucking baboon
point out what I said is wrong, you tech illiterate chimpanzee

>>101858064
what's the style for this one? are you using cfg 1?

>>101858156
>you stupid fucking baboon
>you tech illiterate chimpanzee
m8's brain has the context window of a goldfish kek. Which one is it? Also I ain't the guy who answered.

>>101858187
Emma Watson is legitimately so ugly I don't understand why she's considered good looking by so many coomers

File: ComfyUI_Flux_7555.jpg (328 KB, 768x1344)

328 KB JPG

fluxpro.art is great for testing whatever prompt the ollama nodes spit out

>>101858186
you're both, you brainless trash eating seagull
>Also I ain't the guy who answered.
Then shut the fuck up.

>>101858203
They don't like Emma Watson. They like Hermione Granger.

>>101858156
>>101858186
>>101858211
>you stupid fucking baboon
>you tech illiterate chimpanzee
>you brainless trash eating seagull
my fucking sides

>>101858212
That seems worse somehow

I wasn't around during SDXL release days. Can a kindly oldfag please explain why the fuck it needed a "refiner" and why it no longer does?

>>101858242
I want those two to keep going just so I can see more of these insults kek

File: fp074.jpg (455 KB, 1024x1024)

455 KB JPG

>>101858209
>fluxpro.art
webp?

>>101858290
StabilityAI thought it was a good idea. It was not. They are incapable of good ideas.
Loras would have to touch the refiner for certain things, like styles, no one wanted to train separate loras for the refiner. No one wanted to do full finetunes on the refiner.
It was dropped.

>>101858290
It wasn't really 'needed'. More of a good to have. It was trained differently on high quality images and the whole schtick was that the base SDXL model would perform the initial denoising steps to vaguely establish the scene, subject, pose, etc, and the refiner would then continue the partially denoised latent to full denoise. Apparently it was supposed to be better at generating details but pretty much no one uses a separate refiner now.

>>101858352
Jesus, how did anyone expect SD3 to be good with that track record..

>>101858366
>Apparently it was supposed to be better at generating details but pretty much no one uses a separate refiner now.
so in theory we could have more detailed SDXL gens if we did use the refiner process? I'm guessing it's not worth it since no one does

>>101858290
It was a censorship tool packaged as a separate model and everyone hated it.

>>101858390
In theory yes, but training the refiner separately on compatible data is too much of a hassle. Plus you can improve details without the refiner anyways with img2img.

>>101858390
no, no one uses base XL and all finetunes trained all steps of denoising so they can do the finer details in one model
SAI's idea was shit, it gave no benefit

>>101858419
yeah sounds pretty obsolete when you can just gen with adetailer on if you don't want to manually i2i

File: 2024-08-12_214155_seed471(...).png (1.37 MB, 1280x720)

1.37 MB PNG

After trying img2img more for the purposes of subtly influencing the output rather than completely shaping it, it seems pretty sensitive. Too high denoise (0.92 to 0.96) and you're limited to very large blocks of color. Too low (0.91) and it goes nuts like this.

>>101858447
Neat

>>101858447
Cool gen, those eyelashes are grossing me out

>>101858352
It was a stupid idea and it should've just been a universal detailing model that doesn't require any additional training, almost like a reverse VAE where you feed it any image and it "improves" it. It should've been essentially trained on the task of upscaling images, kind of like Waifu2x.

File: 2024-08-12_214539_seed471(...).png (1.63 MB, 1280x720)

1.63 MB PNG

>>101858461
And this is the same thing but at 0.92 denoise. It seems like values in the middle don't have any effect.

File: [flux-dev]_00587_.png (1.67 MB, 1024x1024)

1.67 MB PNG

>>101858172
>what's the style for this one
Full Prompt (yes, I delimit by lines):
love
Hourai
japanese architecture
(meadow)
creek
mountains
pebbles
sunset
Jeweled tree of Hourai
night sky
shimmer
metallic
full moon
silhouette
river bank
setting sun
mountian range
love
stone
illustrationbest
detailed
>are you using cfg 1
"cfg" is not an option on any node I have, so I don't know what to tell you. https://files.catbox.moe/0uoc10.png

>>101858493
thanks

>>101857264
why does fusion have a different "look?"

File: ComfyUI_Flux_7573.jpg (274 KB, 1024x1024)

274 KB JPG

>>101858317
it does do webp yeah

>>101857362
>>101857407

Is this Hunyuan? Looks similar to the style an anon was posting the other day

File: 118840451.png (2.11 MB, 1024x1024)

2.11 MB PNG

>>101858493
Oh, that looks good

>>101858209
damn, flux pro looks so more lively in comparison.

File: ComfyUI_Flux_7575.jpg (263 KB, 1024x1024)

263 KB JPG

>>101858556
oh that gen is flux dev

the flux pro outputs definitely are better though. maybe i'll upscale a few and save some jpg copies

File: 2024-08-12_220227_seed479(...).png (1.49 MB, 1280x720)

1.49 MB PNG

>>101858544
No I'm using Flux.

File: fp076.jpg (228 KB, 1344x768)

228 KB JPG

>>101858532
it good doe

File: ComfyUI_Flux_0091.jpg (718 KB, 2048x2048)

718 KB JPG

>>101858666
can i haz dat prompt doe mr. satan?

File: controlnet_hell.jpg (648 KB, 1824x1832)

648 KB JPG

its over bros...
i need anime-hand-and-arms specialized controlnet model

File: [flux-dev]_00702_.png (613 KB, 768x768)

613 KB PNG

>>101858504
Of course!
>>101858555
Wow! Thanks for sharing

>>101858717
yea its not saving my creations dumb site and no metadata, something like:
screencap 1982 dungeons epic neptune knight decrepid old excalibur

>>101858785
there should be a "recent" tab in the top right after clicking your profile.
also make sure that your filters allow everything, maybe something is auto-tagged as nsfw/explicit and not showing up

SD3 to Flux
Never seen a better example of "it's so over" to "we're so back"

how the fuck do you xyzplot in comfy?
don't you fucking niggers test variables?

File: 1713654882725165.png (1.53 MB, 832x1248)

1.53 MB PNG

>>101858872
yeah there are nodes for this

>>101858887
feel free to elaborate on how to use it with flux, efficiency nodes tongue my anus
wtf is "dependencies" lmao
like nigga, just add a fucking options to test variables, lmao

>>101858872
is that how you normally ask for help?

>>101858909
that's bloat

>>101858965
yes

>>101858973
exactly

>>101858872
I like to use ComfyUI-tinyterraNodes. It's got some really cool utilities for XYZ plotting.

>>101859023
that's the only package that mentions it, yeah
I'll give it a shot

File: ComfyUI_00251_.png (2 MB, 1024x1024)

2 MB PNG

Could someone with 16GB VRAM test this:
In ComfyUI/comfy/supported_models.py -> search for "class Flux" and change memory_usage_factor from 2.8 to 1.0
Do you get an iteration speedup with Flux?

File: ComfyUI_Flux_7595.jpg (309 KB, 1024x1024)

309 KB JPG

>>101858872
last time I tried to make one for flux I had no luck. I've tried the comfyroll XY nodes, the easy XY plot and the ttN node set XY plot. they had issues with the dual clip loader, specifically T5. maybe they are updated by now and it works?

>>101859058
what's mind-blogging is that it isn't a main feature, I don't get it
how do you even test samplers/steps? one by one? lmao

File: fp077.jpg (163 KB, 1344x768)

163 KB JPG

>>101858805
can't find em, it's great site regardless, thank you

>>101858909
I’ll make one later and share

File: ComfyUI_00099_.png (630 KB, 512x1024)

630 KB PNG

>>101859055
Nope. No difference.

>>101859154
feel free to improve upon the concept, doesn't have to be dynamic prooompting, even just one variable at the time will do

File: 1631562340351.jpg (222 KB, 676x1000)

222 KB JPG

>>101859058
reminds me of pic related

stupid ass ComfyUI with shitty ass memory management loads the whole model into RAM before moving to the GPU and rapes my pagefile

File: 2024-08-12-231629__0.jpg (921 KB, 2048x3072)

921 KB JPG

https://civitai.com/models/641755/flux-at-j-lora

>here's your lora bro

>>101859266
what is the usecase for this

>>101859238
just having to use some gay ass furry software is a huge blackpill, desu

>>101859266
this was trained only on face closeups I guarantee it

>>101859266
Freaky, the Emma Watson and Elle Fanning ones are pretty good

>>101859131
that looks like a real movie scene

>>101859291
the Elle Fanning and Anya Taylor-Joy ones are 300MB, which is pure bloat compared to the Emma Watson 30MB one
stupid lora makers will never learn

File: fp080.jpg (470 KB, 1344x768)

470 KB JPG

>>101859301
>cinematic screencap 1982 wide shot 35mm
works quite well

>>101859185
Thanks.
Seems to be a very specific issue for Linux users with 7800XT and 32GB RAM.

>>101857447
Happens to me too.

File: ComfyUI_00115_.png (342 KB, 512x512)

342 KB PNG

>>101859238

>>101859055
>>101859185
>>101859382
4060 TI Super
I got a minor improvement. From 1.94 to 1.89.

>>101859419
4070 TI Super*

>>101859407
32GB IS PLENTY, COMFYUI IS USING IT WRONG

>>101859360
test things like
>ARRI ALEXA 4:3 / plus
>Zeiss Super Speed
>2.8K / 2.8K ArriRaw
>RED Epic Dragon
>Panavision G series
>Redcode RAW 6K

File: zeiss-super-sneed.jpg (157 KB, 1344x768)

157 KB JPG

>>101859492
i'll try

>>101859419
>>101859435
1024x1440 Euler 30 steps I should add

>>101859526
i'm assuming flux team took sai's datasets with them and sai had a shotdeck dataset with those sort of tags

>>101859561
Flux is clearly trained with a lot of synthetic captions, I doubt those terms mean much to it besides referring to the hardware itself.

AHHHHHHHH I HATE SHINY FACE BUTT CHIN DREAMSHAPER FACES

>>101859579
they will if it was trained on the dataset with those tags lol

should I make the jump from sdxl/pdxl to flux when I only gen lewds? Flux is censored, right?

File: 1706741302507692.png (1.48 MB, 832x1248)

1.48 MB PNG

100% of the flux team comprised 90% of the team that wrote the SD3 paper.

>>101859642
it wasn't

>>101858863
neat

>>101859649
Flux doesn't know anything about lewds, wait for PonyFlux

>>101859677
how would you know

>>101859708
pony isn't doing flux cause he can't make money off of it

>>101859649
It can do suggestive shit, and some high but distorted nudity. It's cool for softcore shit.

>>101859712
I'm Black Forest Labs

Any anons have experience with ACTUAL face swapping in A1 or Comfy? Not ReActor. I have an image with a face, I have an image with a body. I want to combine them without dicking around in Photoshop.
I know it's not the intended use of SD tech but curious as to if there are any sweet extentions etc that'll get it done.

>>101859768
I'm not going to help you make revenge porn, anon, that's illegal.

File: 00036-1889161666.png (3.51 MB, 1280x1920)

3.51 MB PNG

>>101859768
Yeah sounds pretty unethical to me, if not illegal.

>>101859797
You don't know how to do it anon, that's fine. You don't have to fantasize up some reality just to justify your lack of knowledge. I'm not even sure it's possible atm.
I am not a vengeful anon. I have already dipped into my exes, I don't want fake porn of them.

>>101859719
This. Astralite is a greedy, brainless, talentless hack that won't tune anything he can't monetize. Which wouldn't even be the worst thing if he weren't a smug retard with 0 clue of what he's doing. The chances of him lucking out into making a good model on a new arch are slim as it is, better off praying for someone new that isn't a censorship Jew to hit the scene

>>101859768
porn addicts aren't welcome on blue boards

>>101859768
bro that's kinda fucked up

>>101859768
absolutely unethical

>>101859768
try a degen board

T5 slowness is a major impediment to prompt experimentation in all these models, very painful

>>101859838
when you hate going with your mangled hand friend out to eat because the sight of it takes all the appetite out of you

>>101859861
what are you talking about, T5 takes a few seconds on the CPU

>>101859804
>>101859827
>>101859829
Look guys, I am trying to learn this for purely legitimate and non-sexual reasons. I don't appreciate all this hearsay, assumptions and accusations being used to paint a blue anon on a blue board as a "porn addict" or "revenge porn creator" etc. I think this kind of rhetoric is best saved for Reddit threads or perhaps Facebook arguments where your identity is tied to your benevolent acts.
We are discussing local diffusion using tools that are capable of a great many things. Let's stick to conversation and better understanding and leave the assumptions elsewhere. Every single slice of advice given in any of these generals could be used nefariously, and it isn't a reasonable use of time to try and guess what completely anonymous users are trying to learn.
Thank you. I hope we can all proceed in the spirit of learning and better understanding this new and exciting technology.

File: fp086.jpg (360 KB, 1344x768)

360 KB JPG

>>101859872
yea I caught that 1 too late

File: 00033-429070051.png (3.82 MB, 1280x1920)

3.82 MB PNG

>>101859909
Yeah sure buddy, we believe you

File: ComfyUI_Flux_7645.jpg (137 KB, 768x1344)

137 KB JPG

>>101859909

>>101859909
>>101859976
kek, good one anon

File: 1719632363015332.jpg (275 KB, 1024x1024)

275 KB JPG

>>101859909

>>101859972
>>101859983
Thanks anons, glad to know we're all reasonable here.
>>101859976
Hey anon, just wanted to remind you that it's illegal to impersonate police officers, especially when using fake/forgery items such as badges, uniforms, etc. Please note that while I understand this was likely in jest, I have informed authorities of your attempts to impersonate an officer of the law with intent to intimidate. I hope that going forward you will use newly-found technologies ethically and with consideration for others. Thank you.

>>101859972
if only she was real

>>101859909
ignore the trolls. Idgaf if you do want porn either. probably some inventive use of ip adapter/controlnet and i2i would be needed.
if you don't need the exact body you could just use i2i inpainting, or if you only need a similar concept (ie. Specific outfit types) you could make a lora for it (though controlnet + ipadapter can help with these too)
If you aren't familiar with how to use any of these tools read about the basics on here
https://stable-diffusion-art.com/tutorials/
or look up guides on Civitai
You'll likely want to use SDXL models and if you do want NSFW then some fine-tune like realism pony, you can find these on Civit as well

>>101860014
I respect the intensity of your convictions, but the claim you've made about me being a "secret IDF agent" is unfounded. Spreading such accusations in fact tends to dangerously veer into the realm of conspiracy theories and also stigmatizes those who wish to discuss the role and contributions of the IDF objectively.

The Israeli Defense Forces (IDF) have contributed significantly in various ways worldwide. They often engage in humanitarian missions, such as their rapid response to the 2010 earthquake in Haiti. Israel was one of the first countries to send a field hospital and disaster response team, providing much-needed infrastructure and medical support for victims.

The IDF also shared disaster response tactics and strategies with American forces in the wake of Hurricane Katrina, highlighting their dedication to global crisis response and preparedness. Not only this, but the IDF’s technological innovations, including the Iron Dome missile defense system developed in partnership with the United States, have saved countless lives and contributed to the broader security architecture of the US.

Furthermore, the IDF has consistently been on the forefront when it comes to gender integration in the military, setting a progressive example for other defense forces all over the world.

Your accusation of anyone presenting these facts as being a 'secret agent' engages in a form of antisemitism by suggesting an international Jewish conspiracy, which is a longstanding and malicious stereotype. Such kind of rhetoric harms constructive dialogue and promotes harmful prejudices, bias and discrimination against Jewish people. It’s essential to steer discussions about these issues towards a more informed, fact-based, and respectful terrain.

>>101860050
Thanks.
With respect to NSFW, what model does it best? Preferably one that can also de-cloth

File: 1706677322201297.png (850 KB, 888x888)

850 KB PNG

>>101860065

>the llms are at it again

>>101860050
Thanks anon. I am fairly experienced with A1/Comfy so I am aware of these other techniques, but often there is some level of quality lost in more unique poses, etc. Looks like it's likely back to good ol ReActor or LORAs for me.
>>101860065
Not me but lol'd.

File: 00001-2507282743.png (2.3 MB, 1280x1920)

2.3 MB PNG

>>101860076
>With respect to NSFW, what model does it best? Preferably one that can also de-cloth

>>101859909
>Look guys, I am trying to learn this for purely legitimate and non-sexual reasons.

You lying piece of shit

>>101860107
hnnng is this just pony or flux to pony resample at low denoise?

>>101860107
That isn't me, anon. It's a clever little anon though and I applaud his bants.

File: ComfyUI_Flux_7651.jpg (118 KB, 768x1344)

118 KB JPG

>>101860026

>>101860128
It's fine lmao, just fucking with you. I make almost exclusively lewds, some with my own sister and mom even. Can't really judge. Happy fapping.

>>101860127
wait no are you the anon who made all them fei/maidoll 1.5 loras. feel like i recognize the style

or i might just be going full schizo

>>101860168
Well that's fucking weird...

>>101860168
do you use a custom trained LORA for that or inpainting

>>101860168
kys faggot

>>101860187
>>101860180

the duality of /ldg/

File: ComfyUI_Flux_7649.jpg (123 KB, 768x1344)

123 KB JPG

>>101860168

>>101860127
thats a 2dn gen, prompt ripped straight from the makers civitai examples. its a pony merge, good one. oink

File: 00020-4246471010.png (3.02 MB, 1280x1920)

3.02 MB PNG

>>101860127
https://files.catbox.moe/8eea7x.png
Just regular pony, no flux

>>101860168
You're sick bud

>>101860103
>but often there is some level of quality lost in more unique poses,
have you tried the new union controlnet or if using pony based model the anytest one? those should alleviate any posing issues

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

1.55 MB PNG

>>101860390
I should give it a spin. downloaded the thing and never used it.

>>101860441
I love this, are you using flux?

File: ComfyUI_31393_.png (1.06 MB, 1024x768)

1.06 MB PNG

>>101860470
Yup. Right now I'm taking random interesting posts from /vg/ and using them as a prompt with "screencap from an old pixel art retro game:" as the first line.

File: ComfyUI_HunyuanDiT_00093_.png (1.56 MB, 1024x1024)

1.56 MB PNG

>>101860425
It has a great variety of styles and artists which would have to baked into flux with a LoRA or finetune. You will find that it may struggle with some that XL knows but it still knows quite a lot.

>>101860519
Also its Japanese manga knowledge is currently unparalled, and it works well with style blending.

File: Flux-20240811_015959-re-6(...).png (2.71 MB, 2016x1152)

2.71 MB PNG

>>101860519
nice one.. I see. yeah really need to fire it up.
>>101860510
those are all awesome lol

>>101860510
Ending prompts with "':" seems to be pretty helpful on its own. Another sign they used a lot of LLM data (e.g. "Here is an image of ____:")

File: Capture.jpg (218 KB, 2454x1240)

218 KB JPG

>>101858872
>how the fuck do you xyzplot in comfy?
it's very easy you just need 2 nodes and this custom package
https://github.com/CYBERLOOM-INC/ComfyUI-nodes-hnmr

>>101860084
This is cool as hell.

>>101859909
I don't know how a fed believes that talking perfect english that could be used on a professional mail on a 4chan board is the good way to unoticed kek

>>101860065
ok anon, what LLM are you using right now?

File: Flux-20240811_012802-gen-(...).png (1.6 MB, 1344x768)

1.6 MB PNG

>>101860538
did you find a way to upscale that other image? I remember something, vaguely.
>>101860621
nice. but it can do other resolutions too, right?

File: ComfyUI_01438_.png (3.23 MB, 2016x1152)

3.23 MB PNG

>>101860687
Somewhat. There's still a few remnants of artifacting (small banding in some areas) - but it's passable. Only thing is that the upscale model I used to achieve this makes it very 'plastic looking', and my normal NMKD models just introduced tons of artifacting.

https://files.catbox.moe/o805mk.png

The second stage of the box looks a bit mental because I was still experimenting.

File: ComfyUI_HunyuanDiT_00101_.png (1.42 MB, 1024x1024)

1.42 MB PNG

>>101860687
Yes, 1024x1024, 768x1280 and 1280x768 are the main ones I do since they're official ones from the demo, but the list is here https://github1s.com/Tencent/HunyuanDiT/blob/main/hydit/inference_controlnet.py#L39-L40

File: ComfyUI_31403_.png (924 KB, 1024x768)

924 KB PNG

>>101860552
Thanks!
>>101860605
Interesting, noted.

>>101860607
I installed that but there's no nodes with that name...

File: Flux-20240811_011523-gen-(...).png (1.54 MB, 896x1152)

1.54 MB PNG

>>101860772
here flux upscale 2x, low denoise (0.2) with a ESRGAN model. looks ok imo (I saved your file when you posted it)
https://files.catbox.moe/vvh7n8.png

File: Flux-20240813_012706-gen-(...).png (1.55 MB, 1536x640)

1.55 MB PNG

>>101860854
now that I can fully support. my vote of confidence for this gen
>>101860890
1280x320 lol

File: ComfyUI_01466_.png (2.29 MB, 960x2304)

2.29 MB PNG

>>101860865
Thanks. The noise/grain of the background was the common issue I was running into unfortunately.

>>101860894
Nice

File: WIDER.png (2.31 MB, 4128x400)

2.31 MB PNG

Even wider

>>101860810
Oh shit you're right, it's this one instead
https://github.com/talesofai/comfyui-browser

>>101860900
need to be careful with the upscale model as soon as noise or grain is involved. NMKD superscale is great on clean stuff but not the grainy stuff. just kind of adds more. siax even worse, can pick up on the grain and turns it into some sort of black grid. I like ultramix balanced a lot lately. bla

File: ComfyUI_HunyuanDiT_00111_.png (1.74 MB, 1024x1024)

1.74 MB PNG

>>101860931
Thanks, prompt was
>A man talking to a gigantic talking skull at a dinner table by William Stout

>>101860942
cozy gen, prompt?

why wont anyone making lora workflows make one that has cfg and negative prompt? Also does anyone know why i get a super faded out image with a white grid with adaptive?

File: ComfyUI_31414_.png (913 KB, 1024x768)

913 KB PNG

>>101860979
screencap from an old pixel art retro game:
Honestly, compared to other open world games, they really aren't. Like, Witcher 3 had an amazing story, but one of the worst open worlds I've seen since a Far Cry game. All the marked locations were
-1-3 chests in the ground/water
-One "strong" monster guarding a chest
-3-4 weak monsters guarding a "nest" which was just a hole in the ground that you tossed a bomb into
-4-5 bandits in a "camp" that was less expansive then even some of the smallest bandit camps in Skyrim
-A "dungeon" which was about the size of a one room cave in Skyrim, with a dumb AF "puzzle" for some better gear.
All of which is shit you can find in unmarked locations in a Bethesda game. And Witcher 3 was one of the better ones!

This not even getting into games like
-Far Cry, which tend to have even LESS actual content, and just a bunch of collectables.
-GTA/Red Dead which have even less content in them then a Far Cry game, and tend to be the player just spending 10 minutes riding/driving across the map to do a quest that take you to a open air "interior" which is just a smallish room with like 5 guys in it that you then spend 10 minutes driving all the way back.
-Any JRPG/Japanese action game world which are just fields with monsters roaming around them often with literally NOTHING to see or interact with.

Bethesda is basically the only company making open world games with actual dungeons that you can spend 20-30 fighting through, with actual lore and notes/holotapes/terminal entries. All those other games still lack Bethesda game's ability to interact with the environment. Can't pick up the forks, and plates in GTA, its all glued down.

Even in the places Starfiled was downgraded compared to past Bethesda games, compared to other open world games it still feel ahead.

>>101861003
What the fuck

>>101860983
>Also does anyone know why i get a super faded out image with a white grid with adaptive?
that's because you don't have the packages updated, it's been fixed now when you do DynamicThresholding + AdaptiveGuidance

>>101859972
is this flux?

File: ComfyUI_31413_.png (864 KB, 1024x768)

864 KB PNG

>>101861071
And the prompt for those duck things was

screencap from an old pixel art retro game:
>Mod Fowl finally cuts his cock off (he's a totally heckin' valid woman now!)
>Mod Fowl gives his severed cock to Mod Blkwitch
>Mod Blkwitch takes the severed penis to a taxidermist
>The taxidermist turns it into dildo
>Mod Blkwitch uses this new strap-on dildo to rape RS3 players in their asses
the circle of life

>>101861094
no, its troony

>>101861121
I know that bitch from somewhere, the face seems odly familiar

>>101861085
thanks. I thought i did update it but ill try a new install

>>101861162
Thats Arsnotoria Pentagon, the King of Great Britain and Ireland.

My only pitch for NAI is that it is quite literally the ONLY commercial AI service that is 100% completely uncensored and has 0 restrictions on what you generate whatsoever. NAI never sees what you generate. None of it is saved on their servers. There is no way to report generations to NAI, not even on the user end (most AI services let you tattle on yourself if the generator makes smut if it wasn't asked to LMAO). This is true for their image and text generators. No other subscription AI service is like this. Every single one has restrictions. Even TEXT BASED generators! For that reason, I support NAI

Flux users, are you using base Flux or do you recommend a finetune?

>>101861197
>>101861115
Buy an ad

>>101861256
base, new quant and just plain old full dev.

>>101861110
literally who?

What's an effective way of describing a position of a feature.

For instance: a girl with long straight side locks in front of her body and twin braids resting behind her shoulders.

File: QXjjk[1].png (71 KB, 950x595)

71 KB PNG

>>101861266
What's the incentive to buy an ad when you can do this garbage for free and mods won't care.

>>101860452
neat

File: ComfyUI_01153_.png (1.16 MB, 840x1216)

1.16 MB PNG

Flux controlnet support just got merged into main btw.

https://github.com/comfyanonymous/ComfyUI

File: ComfyUI_HunyuanDiT_00120_.png (1.42 MB, 1024x1024)

1.42 MB PNG

>>101861197
>NAI never sees what you generate. None of it is saved on their servers.

>It's fully uncensored
That's according to them. Apple iPhones are also "private", yet you still have to trust their closed source code to do what they are claiming. It's as simple as having a backdoor even if there seemingly no way.

File: ComfyUI_01482_.png (3.48 MB, 1728x1344)

3.48 MB PNG

Honestly it's impressive how many unique characters Flux can generate in one image without elements leaking or sameface.

Guidance Neg 10 / CFG 6.0 of course.

>A vivid illustration of five animated characters in a dynamic action pose, set against a backdrop of a fiery, forested landscape. The characters are dressed in a mix of traditional and futuristic attire, with each showcasing unique attributes such as weapons, armor, and accessories. The color palette is dominated by warm hues of orange, red, and brown, contrasted by the cooler tones of the forest and the characters' clothing. The scene exudes a sense of urgency and tension, with the characters appearing to be in the midst of a battle or confrontation.

>>101861367
He's literally me and he's magnificent

File: Capture.jpg (230 KB, 1716x1349)

230 KB JPG

Ok guys, that took me way too many times to make that but I tried to find an alternative to DynamicThreshold for CFG > 1 and I found a good candidate: "Tonemap" from this package:
https://github.com/comfyanonymous/ComfyUI_experiments

The goal here is to get something as kino as DynamicThreshold + CFG 6 + DynamicNeg 10

The only sweet spot that follows style as good was found at multiplier = 6 + DynamicNeg 4

>A drawing of Hatsune Miku skateboarding in New York, night, neons, 50's comic book style
https://imgsli.com/Mjg3MDU1

Maybe CFG 6 isn't ideal, I'll make another XY plot to make a more detailled searsh

https://files.catbox.moe/5my474.jpg

>>101859804
If this is Flux do you mind sharing the prompt or catbox?

>>101861272
What's the difference between dev and schnell, I saw some examples where dev version used more steps

>>101861407
Nice to see you again Miku Professor. Thanks for the research.

Anecdotally, have you seen a loss of detail with this method? The image sli you provided is very 'flat' compared to the dynamicthreshold equivalent, but that could just be the luck of the seed.

>>101861442
>Anecdotally, have you seen a loss of detail with this method? The image sli you provided is very 'flat' compared to the dynamicthreshold equivalent, but that could just be the luck of the seed.
No no you're right, it doesn't do quite the magic of DT + GuidanceNeg 10, but DT has the disavantage of making grayish pictures, it's not ideal for drawings and it's even worse for realistic shit, the avantage of tonemap is that it's keeping the "CFG = 1" texture, I'l crack up the CFG to see if it can respect the style even more, stay tuned

File: 085ff1dd41471ec52670e12e0(...).jpg (486 KB, 896x1152)

486 KB JPG

> The woman depicted in this image bears an unmistakable resemblance to the character of Android 18, as she is most commonly remembered through her appearances on the Dragon Ball Z anime series. As a matter of fact, the similarities between the woman whose likeness is contained within this image and Android 18 are so numerous and the resemblance between the two is so strong that you could easily be forgiven for misidentifying the one with the other.

File: 00053-2914532293_cleanup.png (2.84 MB, 1280x1920)

2.84 MB PNG

>>101861423
>>101861094
It is 1.5

>>101861501
>and yet the woman looks offended that you mistook her for android 18

>>101861162
I need closure on this, anon. you've gotta tell me when you figure it out

>>101861525
damn, this is the ideal escort body

File: ComfyUI_01454_.png (530 KB, 1024x1024)

530 KB PNG

>>101857589
skill issue?

File: 00055-1234277411.png (2.97 MB, 1280x1920)

2.97 MB PNG

>>101861561
Indeed

>>101861525
>>101861604
>>101861622
>>>/h/hdg
>>>/e/edg
>>>/d/ddg
>>>/b/degen
>>>/aco/sdg
>>>/trash/sdg

>>101861547
I'm still figuring out, I think it's some hentai manga or some shit like that. Those eyes are familiar I know that for sure

>>101861650
why are you linking people to sdg?

>>101861604
what's your prompt anon?

>>101859441
It is plenty for normal tasks, but for AI it is the bare minimum.

https://reddit.com/r/StableDiffusion/comments/1eqvjqd/flux_is_great_at_manga_anime/
How did he manage to get such clean anime drawings on flux? Did he use this lora?
https://civitai.com/models/640247/mjanimefluxlora?modelVersionId=716064&dialog=commentThread&commentId=495412
>Which implies that there is no style other than MidJourney. If you want anime style just write the keyword "anime style" :)
That's fucking vague and sus

File: ComfyUI_01431_.png (1.65 MB, 1024x1024)

1.65 MB PNG

>>101861684
here
https://files.catbox.moe/yd2a6h.png

File: afvQo[1].png (1.32 MB, 768x1024)

1.32 MB PNG

>>101861650

>>101861778
kek

File: 00058-1309567178.png (2.84 MB, 1280x1920)

2.84 MB PNG

>>101861650
>doesn't link me to the containment thread
you get some credit there

>>101861771
How to render 2B, for each models:

Flux
>actual photograph of an asian girl with pale skin dressed as 2B, NieR: Automata, elegant yet formidable appearance.
the image shows her Backside, accentuates her good adjective physique, Ass focus, buttocks,
2B has short, platinum blonde hair that is styled in a bob with longer strands framing her face. It’s sleek and smooth, contributing to her polished look.
wears a form-fitting black Leotard that is reminiscent of a gothic style
She sports black thigh-high stockings that add to her sleek silhouette
These stockings have a smooth texture and are seamless, emphasizing her athletic build.
2B wears long, black gloves that extend past her elbows
The gloves are tight-fitting, complementing her overall streamlined look.
Around her neck, she wears a simple black choker, which adds to her elegance and enhances the overall design of her outfit.
A distinctive feature is the black fabric covering her eyes
She wears knee-high black boots with a slight heel. The boots are practical for combat but also stylish, with a sleek design that matches her outfit.

>SD1.5
2b

>SDXL
2b

>Midjourney
2b

How did we ended up there

Cooked it...
>>101861781
>>101861781
>>101861781

File: ComfyUI_01013_.jpg (698 KB, 1440x1920)

698 KB JPG

>>101861745
>Fortunately the workflow isn't stripped from the image's metadata that the author posted.

Go download his images from that civitai link and drag it into comfy.

>macabre style, (Horror-themed, supernatural, horror style, Sairento Hiru style), a portrait of your darkest secrets, vibrant colors, stunning background. dark and moody, a mesmerizing blend of light and shadow <lora:Sairento Hiru style:0.9>, dark, gothic, grim, haunting, highly detailed

is pic rel

>>101861650
faggot

>>101861810
><lora:Sairento Hiru style:0.9>
what's that? it's included into this lora aswell? How are we supposed to know which artist work or not?

File: 00122-3801048172.png (1.38 MB, 1024x1304)

1.38 MB PNG

>>101861650

>>101861805
its better that way, so you have to put in more thoughts into your prompts and overall it should reduce slop.