[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: album.png (1.35 MB, 1024x1024)
1.35 MB
1.35 MB PNG
Discussion of free and open source text-to-image models

Previous /ldg/ bread : >>101787894

>Beginner UI
EasyDiffusion: https://easydiffusion.github.io
Fooocus: https://github.com/lllyasviel/fooocus
Metastable: https://metastable.studio

>Advanced UI
Automatic1111: https://github.com/automatic1111/stable-diffusion-webui
ComfyUI: https://github.com/comfyanonymous/ComfyUI
InvokeAI: https://github.com/invoke-ai/InvokeAI
SD.Next: https://github.com/vladmandic/automatic
SwarmUI: https://github.com/mcmonkeyprojects/SwarmUI

>Use a VAE if your images look washed out
https://rentry.org/sdvae

>Model Ranking
https://imgsys.org/rankings

>Models, LoRAs & training
https://civitai.com
https://huggingface.co
https://aitracker.art
https://github.com/Nerogar/OneTrainer
https://github.com/derrian-distro/LoRA_Easy_Training_Scripts

>Flux
https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnell
https://comfyanonymous.github.io/ComfyUI_examples/flux

>Pixart Sigma & Hunyuan DIT
https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma
https://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiT
https://huggingface.co/comfyanonymous/hunyuan_dit_comfyui
Nodes: https://github.com/city96/ComfyUI_ExtraModels

>Kolors
https://gokaygokay-kolors.hf.space
Nodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper

>AuraFlow
https://fal.ai/models/fal-ai/aura-flow
https://huggingface.co/fal/AuraFlows

>Index of guides and other tools
https://rentry.org/sdg-link
https://rentry.org/rentrysd

>GPU performance
https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html
https://docs.getgrist.com/3mjouqRSdkBY/sdperformance

>Try online without registration
txt2img: https://www.mage.space
img2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest
sd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium

>Related boards
>>>/h/hdg
>>>/e/edg
>>>/d/ddg
>>>/b/degen
>>>/vt/vtai
>>>/aco/sdg
>>>/trash/sdg
>>
File: 1722484139767935.jpg (331 KB, 1536x2048)
331 KB
331 KB JPG
>mfw
>>
File: ComfyUI_01751_.png (1.9 MB, 1024x1024)
1.9 MB
1.9 MB PNG
For those who weren't on the previous thread, I accidently found a way to remove flux's bias towards generic styles and Miku's overcooking, increase the GuidanceNegative, here's the workflow: https://files.catbox.moe/xlxd00.png

https://imgsli.com/Mjg1Nzc0
>>
File: FD_00087_.png (836 KB, 1024x768)
836 KB
836 KB PNG
>>
WHEN THE FUCK WILL FLUX BE ABLE TO MAKE PORN??!! HELP!!!!
>>
>>101793249
2 weeks
>>
>>101793249
in two weeks
>>
>>101793249
It doesn't make porn? sorry, my bad. I will fix the code base tomorrow. I uploaded wong model
>>
>>101793249
Just reinstall sdxl anon, we are months if years away from that.
>>
>>101793249
can you make flux schnell but 5bit?
>>
File: FD_00139_.png (802 KB, 1024x768)
802 KB
802 KB PNG
>>101793329
I think someone is already doing that. I remember someone posted a tweet
>>
File: ComfyUI_temp_rrpzr_00095_.png (1.53 MB, 1024x1144)
1.53 MB
1.53 MB PNG
>>
>>101792935
Flux can't into deformity, so far, but maybe...
>>
File: ComfyUI_00014_.png (1.36 MB, 1344x768)
1.36 MB
1.36 MB PNG
>>101793369
thanks for the censorship, now i can print this out and hang it up at work since it's sfw
>>
File: ComfyUI_01768_.png (2.06 MB, 1024x1024)
2.06 MB
2.06 MB PNG
>>101793059
Finally, Flux fucking listen to my drawing styles
>Hatsune Miku, 80's anime drawing style
https://imgsli.com/Mjg1Nzg2
>>
>>101793359
thanks mr trump I hope you get to be a porn actor with your friend harris
>>
>>101793249
>>
>>101793421
okay I'm sorry
>>
File: file.png (1.68 MB, 832x1216)
1.68 MB
1.68 MB PNG
>>101793059
thanks goblin-anon.

here, for your troubles
>>
File: FD_00144_.png (773 KB, 1024x768)
773 KB
773 KB PNG
>>101793469
Now this is what I expect when I think of a sexy goblina
>>
File: FD_00153_.png (1.18 MB, 1024x1024)
1.18 MB
1.18 MB PNG
>>
File: F_00328_.png (1.34 MB, 1024x1024)
1.34 MB
1.34 MB PNG
>>
File: ComfyUI_00016_.png (1.36 MB, 1344x768)
1.36 MB
1.36 MB PNG
>>
>>101793540
how
>>
>>101793539
>schizomeds is a box of tissues
accurate desu
>>
File: 1713393145733323.png (77 KB, 640x426)
77 KB
77 KB PNG
>>101792978
>op pic
what's LD has to do with Pussia?
>>
>>101793454
I was just fucking with you
>>
File: FD_00068_.png (1.49 MB, 768x1024)
1.49 MB
1.49 MB PNG
>>101793555
>>
Flux's style swings more than pony
>>
>>101793573
well, it is a blue board
>>
File: 456456456484.png (100 KB, 1482x547)
100 KB
100 KB PNG
>>101793059
Can you give me a simple workflow without bloat?
>>
>>101793625
just remove the red nodes they are useless for you
>>
>>101793625
literally just remove those nodes anon and relink
>>
>>101793636
he doesn't even have to relink, if it's deleted between 2 others nodes, the 2 other nodes will link to each other
>>
where is ani?
>>
>>101793625
>I don't like the way you did this, make it again for me
I hate zoomers so much it's unreal
>>
File: ComfyUI_00017_.png (1.29 MB, 1344x768)
1.29 MB
1.29 MB PNG
The duality of trump
>>
>>101793059
did you try going past 10 guidance, anon?
>>
>>101793694
I'm actually trying a XY plot to see where's the sweet spot
>>
>>101793630
>>101793655
top kek
>>
File: FLUX_00030_.png (1.19 MB, 896x1152)
1.19 MB
1.19 MB PNG
>>
File: file.png (2.5 MB, 1024x1024)
2.5 MB
2.5 MB PNG
>>
File: file.png (3.69 MB, 1216x1664)
3.69 MB
3.69 MB PNG
resampling in pony is saving my lack of prooompting skills
>>
File: 1696545545297472.jpg (123 KB, 1024x1024)
123 KB
123 KB JPG
>>101793578
>>
File: ComfyUI_01875_.png (2.22 MB, 1024x1024)
2.22 MB
2.22 MB PNG
>>101793425
>A watercolor ink pen outline painting of Hatsune Miku by Dean Crouser
Watercolor Miku! Still no idea about artist though
>>
File: ComfyUI_00019_.png (1.36 MB, 1344x768)
1.36 MB
1.36 MB PNG
>>
>>101793786
Are the styles constant on multiple seeds?
>>
File: 1694384150643728.png (364 KB, 494x348)
364 KB
364 KB PNG
>>101793540
>>
>>101793814
you can consider 4 different pictures as 4 different seeds?
>>101793786
>>101793425
>>101793059
>>101792754
>>
>>101793786
>Watercolor Miku!
Nice anon, glad it's working on watercolor aswell
>>
>>101793842
can you make multiple gens of the same prompt with different seeds?
I see style swings across the same prompt
>>
File: Capture.jpg (797 KB, 3840x1360)
797 KB
797 KB JPG
>>101793059
Ok I did a XY plot (X = sampler I just took euler AND Y = GuidanceNeg scale) and... what the fuck, it looks like the sweet spot is at 10... how lucky was I? I was just puting this number randomly, if I decided to go for 7 for example I would've gotten something glitched and I would never consider this viable or look at it with more details... this is INSANE
https://files.catbox.moe/rpp8u2.jpg
>>
File: FD_00192_.png (1.29 MB, 1024x1024)
1.29 MB
1.29 MB PNG
>>101793726
OK post progressives
>>
File: ComfyUI_00021_.png (1.57 MB, 1216x832)
1.57 MB
1.57 MB PNG
>>
>>101793916
Ok, let me try 4 random seeds and see what we'll get
>>
>>101793059
can you screenshot or discuss the tweaks? I'm busy genning.
>>
File: 1723163822917093.jpg (1.22 MB, 3223x3051)
1.22 MB
1.22 MB JPG
>>101793970
>can you screenshot or discuss the tweaks? I'm busy genning.
Sure
>>
>>101793940
is it me or is the lowest value the best gen?
maybe try going from 3.5 to some lower values in another plot
>>
>>101794035
>is it me or is the lowest value the best gen?
the lowest value (3.5) doesn't respect the 80's anime style drawing at all, it's the default value that gives you the generic slop flux style

>>101794035
>maybe try going from 3.5 to some lower values in another plot
you have my workflow and it has the XY plot node, you can try it by yourself if you want
>>
>>101794057
>you can try it by yourself if you want
that's exactly what i would do if i wasn't a complete retard.
>>
File: Flux_00540_.png (900 KB, 1024x1024)
900 KB
900 KB PNG
guys I'm demoralized
>>
File: Capture.jpg (199 KB, 2182x1196)
199 KB
199 KB JPG
>>101794084
it's really not that hard anon, you just have 2 nodes to deal with, look at this example and just try to do the same for what you want to do. If one of those nodes are purple it means it's deactivated, to activate it you right click on it -> Bypass
>>
File: ComparaisonDifferentSeeds.jpg (1.75 MB, 3164x6216)
1.75 MB
1.75 MB JPG
>>101793916
>>101793963
X = seed
Y = GuidanceNeg scale
>>
>>101794116
Still inconsistent across seeds, this is a model problem
>>
File: ComfyUI_01801_.png (2.22 MB, 1024x1024)
2.22 MB
2.22 MB PNG
>>101793059
>>101793786
Watercolor works so well with that method
https://imgsli.com/Mjg1ODI5
>>
File: ComfyUI_01890_.png (2.35 MB, 1024x1024)
2.35 MB
2.35 MB PNG
Not bad desu, it recognizes more of Van Gogh's style now
>>
>>101794158
>Still inconsistent across seeds, this is a model problem
can you elaborate? seeds are supposed to give you a completely different picture (in terms of character poses at least)
>>
>>101794191
no one cares about your shitty twitter artist slop
>>
>>101794211
Many models will have a consistent style when given a name of artist and types, Like this is a oil painting of miku but it looks like it's drawn by different people with enough variations to be called a different style.
It's like saying all cats look the same if they have the same general coat coloring
>>
>>101794212
>Van Gogh
>twitter artist
That's a high quality bait right there
>>
>>101794235
https://x.com/vangoghartist

cope harder
>>
File: IMG_2131.png (2.25 MB, 1336x1336)
2.25 MB
2.25 MB PNG
j
>>
>>101794227
>Like this is a oil painting of miku but it looks like it's drawn by different people with enough variations to be called a different style.
Well, the prompt is "a 1700s painting", so of course the style is gonna vary, not all artists from the 1700s painted the same
>>
File: 00000-841294845_cleanup.png (2.65 MB, 1280x1920)
2.65 MB
2.65 MB PNG
>>
>>101794244
kek'ed
>>
>>101794249
i think the point is anon that models are meant to have consistent training data that points to "1700s painting", especially when none of the other variables are changing
>>
File: CatJak_00097_.png (1.73 MB, 1024x1024)
1.73 MB
1.73 MB PNG
>>101794249
Of course but from my test even with a named artist the styles wing greatly
>>
File: FLUX_00036_.png (1.08 MB, 896x1152)
1.08 MB
1.08 MB PNG
>>
>>101794271
that's because the model probably doesn't know the artist? or if it does, maybe it's doing different styles because such artist had different styles to begin with?
>>
>>101794099
checked and okay, i tried to load the catbox but i think it's not showing up in the load files because it's a jpg. there are png files in there i can load but obviously they're not the ones i'm trying to get to. i told you i'm retarded.
>>
>>101794289
It's a weakness in the model itself, there should be a standard when doing such things, every combination should create it's style. This was a heavy criticism of base pony which oddly enough many of the people that made this model's teammates mocked.
>>
>>101793756
How about just put sexy in the negative?
>>
>>101794300
>checked and okay, i tried to load the catbox but i think it's not showing up in the load files because it's a jpg.
how the fuck can it be a jpg? it's literally written .png here >>101793059
>https://files.catbox.moe/xlxd00.png
>>
File: file.png (47 KB, 566x284)
47 KB
47 KB PNG
whoops got the wrong one i guess teehee
>>
>>101794420
kek, that picture was just a comparaison picture, not a workflow
>>
File: Capture.jpg (219 KB, 3501x1383)
219 KB
219 KB JPG
>>101794432
>>101794420
oh yeah I completely forgot btw, you have to install this node on the comfyUi manager to get the XYZ plot node
>>
>>101794423
>>101794423
>>
File: ComfyUI_01121_.png (845 KB, 1024x1024)
845 KB
845 KB PNG
>>
>>101794444
sorry for the dumb question, but how would i go about finding and using new nodes? i'm still just using the comfyui noob quickstart workflow.
>>
File: Capture.jpg (771 KB, 3840x1789)
771 KB
771 KB JPG
>>101794461
1) Install the comfyUi manager
https://www.youtube.com/watch?v=5AHBRVaJf9k

2) Once it's done you'll see a "Brower" button on the Ui, click on it and then click on "Custom Nodes Manager", you'll be able to find the "hnmr" node I was talking about
>>
File: 1713146669243330.jpg (157 KB, 738x1292)
157 KB
157 KB JPG
I created this superhero in my mind long time ago finally got what i imagine his look .
Do you guys have you own OC superheros ?
>>
File: FLUX_00027_.png (1.07 MB, 896x1152)
1.07 MB
1.07 MB PNG
>>101794491
yeah
he's just a dude with a lockpick and a steak knife
>>
>>101794491
Was his superpower having his cock burned off in a horrific accident?
>>
>>101794444
>oh yeah I completely forgot btw, you have to install this node on the comfyUi manager to get the XYZ plot node
What does this mean, and checked

>>101794454
cool product, 10/10 would buy
>>
File: Final_00665_.png (2.46 MB, 1728x1344)
2.46 MB
2.46 MB PNG
>>101794491
i did start creating adventuring-party OCs at one point
>>
>>101794484
thanks, i'll have a looksee
>>
>>101794512
>What does this mean, and checked
it means what it means... if you want to use the XYZ plot node, you have to install the "ComfyUI-nodes-hnmr" node on "Custom Nodes Manager" on the "Browser" menu >>101794484
>>
AHHHHHH

>I updoooted
>now vaedecode(tiled) crashes
>>
File: Capture.jpg (104 KB, 3147x655)
104 KB
104 KB JPG
>>101794512
>>101794517
oh yeah one last thing, you also need the dynamicthresholding node, it's here
>>
>>101794525
what's xyz plot node?
>>
>>101794553
it's to make XY plots, like that one: >>101794116
>>
File: 00003-4205527924.png (2.46 MB, 1280x1920)
2.46 MB
2.46 MB PNG
>>
>>101794503
He just have a big cock.
>>
>>101794484
>>101794548
thanks!
>>
File: FD_00009_.png (1.02 MB, 768x1024)
1.02 MB
1.02 MB PNG
>>101794491
Woman Man. She learned how to read.
>>
File: Flux_00502_.png (1.3 MB, 1024x1024)
1.3 MB
1.3 MB PNG
>>101794491
>Do you guys have you own OC superheros ?
yeah, hes called "Rapeman".
>>
>>101793630
I was warned for this post, meanwhile the guy yesterday freely posting cheese pizza.
Jannies are weird.
>>
>>101794559
>one of my cousins looks like her
Well that's odd.
>>
File: 00006-4103237681.png (2.39 MB, 1280x1920)
2.39 MB
2.39 MB PNG
>>101794691
>I was warned for this post
Rightfully so
>>
>>101794692
female cousins or male cousins? :d
>>
>>101794699
keek
>>
>>101794697
Yes CP is OK?
Understood and kys.
>>
File: Elf.png (1.89 MB, 1280x1024)
1.89 MB
1.89 MB PNG
>>101794697
>>101794585
Nice.
>>
File: CatJak_00129_.png (2.03 MB, 832x1280)
2.03 MB
2.03 MB PNG
>>
>>101794691
One of my satire posts of the Trump ********* was deleted too. I don't think 4chins wants anything to do with it outside of /pol/ probably
>>
>>101794559
Is there a way to get comfyui to unload the model from ram, so the vaedecoder has more ram?
>>
File: CatJak_00131_.png (2.09 MB, 832x1280)
2.09 MB
2.09 MB PNG
>>101794717
same prompt style swing
>>
>>101794719
next time just include a naked child in the image and you will be safe apparently.
>>
File: 00008-4176667973.png (2.85 MB, 1280x1920)
2.85 MB
2.85 MB PNG
>>101794709
I didn't say CP was ok, and the CP posted here gets deleted
>>
>>101794724
you can use this script, it'll have the OverrideMODELDevice node
https://gist.github.com/Sunderbraze/d0b0f942256965b40f54247344fea37f
you put that into ComfyUI\custom_nodes
>>
>>
>>101794737
I stand by the latter half of my comment.
>>
I think t5 is good with typos, or the training data had typos in it
>>
File: FD_00207_.png (2.12 MB, 1024x1024)
2.12 MB
2.12 MB PNG
This dynamic thresholding thing is neat I just wish it didn't take twice as long.
>>
>>101794762
That's the fault of the CFG different to 1, it makes 2 pictures to make it work, that's why it's twice as slow unfortunately
>>
>>101794500
Its just look like guy with ski mask and and skate gear kek
>>101794690
Nice
>>
>>101794761
>I think t5 is good with typos, or the training data had typos in it
I think the opposite, it's really sensitive to typos, at somepoint I wrote "skatbording" instead of "skatebording" and he didn't give me any skate
>>
>>101794778
did it give you skat?
>>
vaedecode(tiled) failed again, so I'm trying the for_testing version.

Then I'll try the unloader.

This is all because I just have 32gb of system ram.
>>
>>101794785
https://en.wikipedia.org/wiki/Skat_(card_game)
kek

fortunately I didn't made a "scat" typo mistake, even though flux wouldn't understand :^)
>>
>>101794741
ty
>>101794787
>>
>>101793059
This doesn't seem to work with schnell.
>>
so glad I bought a gorillion gb of ram to run big LLMs on CPU
>>
File: Capture.jpg (37 KB, 1080x415)
37 KB
37 KB JPG
>>101794787
>>101794793
do you load the fp16 version? I'm using fp8 and there's fp8 files on huggingface: desu I don't see much difference between them and you'll get less painful loads out of it
https://huggingface.co/Kijai/flux-fp8
you load it with the e4 weights aswell
>>
File: Flux_00543_.png (1.04 MB, 1024x1024)
1.04 MB
1.04 MB PNG
>>101794745
I'd give her some VRAM if you know what I mean
>>
>>101794745
too lewd for this board
>>
File: 00012-1599344138.png (2.51 MB, 1280x1920)
2.51 MB
2.51 MB PNG
>>
File: FD_00213_.png (1.81 MB, 1024x1024)
1.81 MB
1.81 MB PNG
>>101794770
Yeah, it's already slow as it is, but I supposes some sacrifices have to be made for quality
>>
>>101794795
another reason to put schnell to trash, dev is clearly the superior version
>>
>>101794817
>the apple magic mouse

kek
>>
>>101794824
nice picture! what prompt anon?
>>
>>101794741
Oh, I meant maybe move the model to swap or something, because vaedecode becomes super slow if it doesn't have enough ram
>>
>>101794850
If I'm being honest with you, just buy more ram, it's automatically puts your model to swap because you don't have enough ram, that's why it's slow
>>
File: FD_00215_.png (2.03 MB, 1024x1024)
2.03 MB
2.03 MB PNG
>>101794842
A Marvel DC Comic book illustration in the style of Jim Lee of Hatsune Miku
>>
>>101794868
thanks anon, I'll add it to my examples there:
https://www.reddit.com/r/StableDiffusion/comments/1enm9og/comment/lh79ucw/?utm_source=share&utm_medium=web2x&context=3
>>
File: Capture.jpg (910 KB, 2295x1698)
910 KB
910 KB JPG
>>101793059
Doing some XY plot between GuidanceNegative (X) and CFG (Y) at the moment, still working... looks like it's starting to work well at CFG + 3 + GuidanceNeg 5
>>
File: 1723174569004921.png (2.5 MB, 832x1280)
2.5 MB
2.5 MB PNG
>>101794727
BMP here, awesome pic/style. Gimped it up a bit for fun, hope you don't mind.
>>
>>101794917
>slurp slurp slurp
>>
File: file.png (3.09 MB, 1248x1824)
3.09 MB
3.09 MB PNG
goodnight anons, have a flux-based pony gen from me
>>
File: CatJak_00136_.png (1.97 MB, 832x1280)
1.97 MB
1.97 MB PNG
>>101794917
Looks better that way, I'm still playing with CGF
>>101794934
Go cry in the corner
>>
>>
>>
https://www.reddit.com/r/StableDiffusion/comments/1emy7uv/comment/lh7e9zl/?utm_source=share&utm_medium=web2x&context=3
>There's an update to the other method that improves performance by using Adaptive Guidance, give it a try!
What? How to find that node?

https://arxiv.org/abs/2312.12487
> Our experiments demonstrate that AG preserves CFG's image quality while reducing computation by 25%. Thus, AG constitutes a plug-and-play alternative to Guidance Distillation, achieving 50% of the speed-ups of the latter while being training-free and retaining the capacity to handle negative prompts.
Holy fuck? It's like a CFG but without the 2x speed decrease? WHERE DO I SIGN?? WHERE IS IT??
>>
>>101794987
>AG preserves CFG's image quality
what really is CFG? And AG?
>>
File: FD_00241_.png (1.55 MB, 1024x1024)
1.55 MB
1.55 MB PNG
>Hatsune Miku, colorful abstract gauge theory particles and fields interacting on a lattice
>>
>>101794997
CFG is something that helps the model understands prompts better, you can see an example here:
https://reddit.com/r/StableDiffusion/comments/1emow5p/finding_the_sweet_spot_between_guidance_and_cfg/
>>
>>101794997
CFG = Cum Fart Gay
AG = Ass Gape
>>
>>101795003
did you use that method anon? >>101793059
>>
File: FD_00234_.png (1.86 MB, 1024x1024)
1.86 MB
1.86 MB PNG
>>101795012
Yes, all my posts for the remainder of this thread at least will be using DT.
>>
>>101795003
nice
>>
File: ComfyUI_00881_.png (2.24 MB, 1024x1024)
2.24 MB
2.24 MB PNG
Pixel Migu
>>
File: download.jpg (7 KB, 298x169)
7 KB
7 KB JPG
>>101795068
Omg it Migu!
>>
File: FD_00252_.png (1.94 MB, 1024x1024)
1.94 MB
1.94 MB PNG
>a Boris Valejo fantasy painting of of Hatsune Miku draped on a rock
>>
>>101794525
>XYZ plot node
can xyz do different samplers?
>>
File: FLUX_00049_.png (1.01 MB, 896x1152)
1.01 MB
1.01 MB PNG
>>
File: Capture.jpg (298 KB, 1783x1745)
298 KB
298 KB JPG
>>101795090
it can do absolutely any nodes that's active on your workflow on that moment, that's why I'm loving that shit
>>
>>101795097
I saw similar at a coffee shop, except she was an employee, and obviously a retard :^)

I could read their minds "don't be racist, she's not retarded, actually, I'm a racist".
>>
>>101795081
Oh wow it looks good!
>>
>>101795081
i've tried like 10 different times with diffreent thnigs to do a boris valejo and had no luck
secret?
>>
Imagine how good a 4k+ flux render looks
>>
File: FD_00253_.png (2.01 MB, 1024x1024)
2.01 MB
2.01 MB PNG
>a Boris Valejo fantasy painting of of Hatsune Miku dressed as sailor moon
By the way reddit guy, the results are consistent
>>
>>101795101
very weird closeby seeds can be similar.
>>
>>101793059
The most impressive part of that method is that to get better styles, you don't need to decrease the GuidancePositive anymore and let it at 3.5 (sweet spot for good anatomy, hands and text)
>>
File: ComfyUI_00886_.png (2.05 MB, 832x1216)
2.05 MB
2.05 MB PNG
Whoever did the flux bias removal is a godsend. Not perfect but it looks less like slop at least
>>
>>101795117
see >>101793059

>>101795120
picrel
>>
File: 00586.jpg (1.78 MB, 2304x1792)
1.78 MB
1.78 MB JPG
>>
File: bob-ross-meme.jpg (49 KB, 400x279)
49 KB
49 KB JPG
>>101795129
>Whoever did the flux bias removal is a godsend.
Thanks anon, it was a complete accident, but what a happy little accident it was! >>101793940
>>
File: FD_00260_.png (2.22 MB, 1024x1024)
2.22 MB
2.22 MB PNG
Holy fuck this one is mental
>>
File: FLUX_00050_.png (1.08 MB, 896x1152)
1.08 MB
1.08 MB PNG
>>
>>101795120
>By the way reddit guy, the results are consistent
Glad to hear that
>>
>>101795137
i always freak out when so much red. gonna try it deletnig all those nodes and reconnecting once this queue ends
>>
File: qcrCofACbgxuYRlbbDW7U.png (950 KB, 1024x768)
950 KB
950 KB PNG
not local but its flux dev
>>
>>101795129
Prompt ?
>>
File: 1715460197973880.jpg (69 KB, 738x1292)
69 KB
69 KB JPG
Good enough
>>
File: file.png (583 KB, 446x644)
583 KB
583 KB PNG
>>101795151
>AAA artstation photobash concept art of Navy seals operator with cyberpunk aesthetic, science fiction with exoskeleton suit and tactical equipment, military realism, highly detailed

I think this one's worth the comparison as well. This is a huge step up for concept art enjoyers
https://imgsli.com/Mjg1ODc0
>>
>>101795162
I tidied the workflow a bit.
Here:
>https://files.catbox.moe/e57ghg.png
>>
File: FLUX_00051_.png (1020 KB, 896x1152)
1020 KB
1020 KB PNG
>>
>>101795209
the one at the left is the one with the flux bias removal right? I think you should redo that imgsli by adding a title (the prompt) and the 2 images captions, something like this:
https://imgsli.com/Mjg1Nzk5
>>
>>101795223
Um that's sexual harassment anon
>>
>>101795129
>>101795137
What's basic here, increase guidance on negative, and set cfg to 10?
>>
File: 00588.jpg (1.3 MB, 2304x1792)
1.3 MB
1.3 MB JPG
>>
File: ComfyUI_00892_.png (2.03 MB, 832x1216)
2.03 MB
2.03 MB PNG
>>101795228
One's from the anon's workflow another is just the default comfy example workflow. I didn't touch the CFG so I think it's at 6.0.

I'm at work so I just did a quick thing. Sorry
>>
>>101795235
Yes but you need to have the node that normalised the cfg back down to 1. That's where the DynamicThresholding comes in.
>>
>>101795209
wtf, the one on the right looks like a plastic toy, the one of the left looks like a Midjourney render, we just unlocked Flux to its full potential there
>>
>>101795215
oh, I see, it's an fp8 workflow.

The dev workflow is kind of weird, with these doubled text boxes.
>>
BMP browses /ldg/?
>>
>>101795250
to get the "comfy example workflow" from that anon's workflow, what you need to do is to put cfg = 1 and right click on DynamicThresholdingFull -> bypass to deactivate that node, in case you're interested
>>
File: FLUX_00052_.png (1001 KB, 896x1152)
1001 KB
1001 KB PNG
>>101795234
she's into it
>>
>>101795255
Yeah I don't like how the default Dev works, it always unloads and re-loads every gen for some reason. The baked FP8 checkpoint is much easier to use imo, but the results do change.
>>
>>101795209
can you give me the image on the right? I need it to redo the imgsli, I'll include it into the examples thank you anon
>>
>>101795260
BMP goes where his masters go
>>
>>101795278
You can stop seething because nobody wants to play uh oh schizo with you
>>
File: 00589.png (1.68 MB, 832x1152)
1.68 MB
1.68 MB PNG
>>
File: ComfyUI_01271_.png (1.6 MB, 1024x1024)
1.6 MB
1.6 MB PNG
>>
File: 00014-577919537.png (2.55 MB, 1280x1920)
2.55 MB
2.55 MB PNG
>>
>>101795235
>What's basic here, increase guidance on negative, and set cfg to 10?
GuidanceNegative on 10 and CFG on 6, and like that anon said >>101795253 you need to download the DynamicThresholding node to make it work
>>
File: CatJak_00141_.png (2.1 MB, 832x1280)
2.1 MB
2.1 MB PNG
It's mocking me
>>
>>101795320
wow, it looks just like the last several hundred!
>>
>>101795333
Cope
>>
>>101795332
nice
>>101795333
cope
>>
File: gpu_paradise.jpg (341 KB, 1024x1024)
341 KB
341 KB JPG
I need decent text prompt for sdxl
>>
Weird that dev is faster. I need more ram so vae is successful, though.
>>
>>101795215
so why no clips?
>>
File: ComfyUI_01274_.png (1.45 MB, 1024x1024)
1.45 MB
1.45 MB PNG
>>
File: FD_00282_.png (1.69 MB, 1024x1024)
1.69 MB
1.69 MB PNG
It does make some things look a bit cooked. I wonder about doing a normal low noise Flux pass over after the gen to see if it can clean it up
>>
>>101795333
Checked.

Look closer, it's a jewish version!
>>
File: Capture.jpg (386 KB, 3222x1437)
386 KB
386 KB JPG
>>101795209
I didn't manage to get the same picture, did you touch anything by chance?
>>
>>101795343
>>101795344
>the discord rushes to defend
lmao the actual state of these fags
>>
>>101795314
>>
File: ComfyUI_01275_.png (1.78 MB, 1024x1024)
1.78 MB
1.78 MB PNG
>>
>nigbo seething /sdg/ is completely dead because of him
sad state of affairs, but he brought it on himself
>>
>>101795357
>It does make some things look a bit cooked.
you can decrease the CFG a bit if you feel it's burning too much, I'm actually doing a XY plot between GuidanceNegative and CFG to find something as good but with a lower CFG
>>
>>101795314
>>101795355
damnit meant this one...another 1080ti enjoyer i see
>>
>>101795375
lol
>>
>>101795380
pretty sure there's only one confirmed nig avatarfag
>>
>>101795353
s/it wise i see no difference picnotrel
>>
File: 00089-2559634564.jpg (1.87 MB, 2048x2432)
1.87 MB
1.87 MB JPG
>>101795371
Only retards that say that are debo and PW.
You have your place just stay there.
You can't out prompt us, you can't have civil conversation with us and you spergs sure as fuck don't improve worth a damn
And when I say us I mean the anons that left the old general to get away from (You)
>>
>>101795153
Nice one anon
>>
>>101795395
get back to your containment thread
>>
>>101795401
>mama's booty
still seething even after your mental health break?
absolutely ngmi
>>
>>101795354
Clip is baked into the model.
>>
>>101795331
Why do you think dev is faster in terms of sits? Like without the fp8 weight dtype.

Just strange.
>>
File: CatJak_00146_.png (1.99 MB, 832x1280)
1.99 MB
1.99 MB PNG
The right cfg value is somewhere out there
>>
>>101795434
>The right cfg value is somewhere out there
elaborate?
>>
File: FD_00002_.png (1.86 MB, 1024x1024)
1.86 MB
1.86 MB PNG
>>101795357
Didn't really do anything. Made it worse if anything. Could just be the gen
>>
How many steps do you guys do with floox?
>>
>>101795439
I have no idea what I'm doing but I found a consistent style finally
>>
>>101795430
>Why do you think dev is faster in terms of sits? Like without the fp8 weight dtype.
faster than what?

>>101795444
that's cool, what values did you find?
>>
>>101795443
20 minimum for dev
>>
File: FD_00290_.png (1.06 MB, 1024x1024)
1.06 MB
1.06 MB PNG
>>101795444
>I have no idea what I'm doing
same, and that's why it's fun.
It's like 1.5 levels of exploration again.
>>
>>101795443
20 steps on dev seems to be working fine, if the text doesn't render well, I increase to 30-40 and it does the trick
>>
>>101795400
Maybe a rocm quirk? I get around 20sits with dev, worse with fp8 weight dtype. then nearly 50 sits with the fp8 checkpoint.

odd.

I don't think my card is getting taxed.
>>
File: CatJak_00137_.png (1.92 MB, 832x1280)
1.92 MB
1.92 MB PNG
>>101795448
Negative guidance to 58
>>
File: ComfyUI_01879_.png (1.93 MB, 1024x1024)
1.93 MB
1.93 MB PNG
>>101795120
I love that one

>>101795459
and CFG still at 6?
>>
File: file.png (2.72 MB, 1024x1024)
2.72 MB
2.72 MB PNG
>>
>>101795468
5
>>
>>101795454
1.5 still makes strange things. It's nowhere near ogre
>>
File: 00001-2781450135.png (3.09 MB, 1280x1920)
3.09 MB
3.09 MB PNG
>>
>>101795120
https://imgsli.com/Mjg1ODc4
Just added your examples to the post, thanks anon
https://www.reddit.com/r/StableDiffusion/comments/1enm9og/comment/lh79ucw/?utm_source=share&utm_medium=web2x&context=3
>>
File: sweat.jpg (267 KB, 1026x1023)
267 KB
267 KB JPG
more sweat
>>
File: ComfyUI_02723_.png (1.02 MB, 896x1152)
1.02 MB
1.02 MB PNG
>>
File: FD_00295_.png (1.64 MB, 1024x1024)
1.64 MB
1.64 MB PNG
>>101795487
Oh we haven't even scratched the surface of this model yet.
>>
File: sweat2.jpg (256 KB, 1052x1056)
256 KB
256 KB JPG
sweat is star I guess
>>
File: FD_00296_.png (1.92 MB, 1024x1024)
1.92 MB
1.92 MB PNG
Even Dynamic Thresholding can't bring back Marilyn it seems. Looks more like Lady Gaga
>>
File: CatJak_00132_.png (2.01 MB, 832x1280)
2.01 MB
2.01 MB PNG
Once I figure it out I can make make this more irl accurate with some embellishments of course
>>
>>101795535
Which is ironic since it's clearly her face in the picture in the background
>>
>>101795535
when the model doesn't know the concept, it just doesn't know, can't do anything about unfortunately
>>
>>101795535
i kinda see it
>>
File: CatJak_00151_.png (1.9 MB, 832x1280)
1.9 MB
1.9 MB PNG
So close to perfect
>>
why does every model that does realism end up with that baseline glossy slop look including flux?
>>
>>101795500
did you discover it or work from what this guy said
https://www.reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/
>>
File: BMP_FLUX_05173_.png (1.26 MB, 648x1200)
1.26 MB
1.26 MB PNG
>>101795260
I always have, just off and on. I also keep up with /v/ Dalle. Never posted here before because my gens were never up to snuff with the quality you guys make. /Sdg/ was kept around for promptlets like me. Although if you want my honest opinion /ldg/ is kind of a worse /DALL-3/ lately but there's some really good gens here as well too.
>>
>>101795576
>did you discover it or work from what this guy said
I'm the same guy for the both of them, look at the leddit username it's the same kek
>>
>>101795567
It's impressive how small of text flux can render
>>101795577
>/DALL-3/
Flux itself is very comparable to Dalle but I agree
>>
>flux bias removal
qrd?
>>
File: mgc.gif (1.39 MB, 275x252)
1.39 MB
1.39 MB GIF
>>101793059
This is actually insane, it was flux's biggest wearkness, it's huge bias on generic slop, and you're telling me by just by adding 10 to Guidance Negative you remove that bias? LOOOOL, it's fucking magic if you ask me.
>>
File: 2024-08-08_00010_.png (1.07 MB, 1280x720)
1.07 MB
1.07 MB PNG
>>101795344
Checked
>>
>>101795599
no more slop on flux anymore with this method -> >>101793059
https://reddit.com/r/StableDiffusion/comments/1enm9og/discovered_by_accident_a_trick_to_make_flux/
>>
File: robotwaifu.jpg (269 KB, 1053x1051)
269 KB
269 KB JPG
Save money for robots.
>>
File: ComfyUI_02727_.png (962 KB, 896x1152)
962 KB
962 KB PNG
>>
File: FD_00002_.png (1.04 MB, 1024x1024)
1.04 MB
1.04 MB PNG
I wonder why this happened
>>
File: Flux_00567_.png (1.26 MB, 1024x1024)
1.26 MB
1.26 MB PNG
>>
>>101795576
oh pls forgive you're doing good. experimentation is the way to find out this shit.
>>
File: 00010-1647997980_cleanup.png (2.93 MB, 1280x1920)
2.93 MB
2.93 MB PNG
>>
File: out.webm (33 KB, 320x157)
33 KB
33 KB WEBM
I asked for some animation frames and it gave me this
I could make a game with this yet, just full AI in all its glory
>>
>>101795656
I walk like this.
>>
>>101795656
Schizo world 2
>>
File: 1699661003039648.jpg (501 KB, 768x1344)
501 KB
501 KB JPG
Sd3 pro can make decent images
>>
>>101795580
ugh this for you >>101795650
i need sleep
>>101795646
how do you get it to put swastikas on the 1080ti?
>>
>>
>>101795680
Did you notice her thumb is on backwards?
>>
>>101795681
>ugh this for you >>101795650
>i need sleep
it's ok lol, and thanks anon :3
>>
File: ComfyUI_00049_.png (1.66 MB, 832x1216)
1.66 MB
1.66 MB PNG
>>
File: 1696787500395529.jpg (264 KB, 768x1344)
264 KB
264 KB JPG
Just prooompt harder
>>
>>101795681
"swastikas on fans", doesnt work everytime tho. sometimes it fucks them up and sometimes it gets them right.
>>
File: FD_00301_.png (1.39 MB, 1024x1024)
1.39 MB
1.39 MB PNG
>>101795567
I don't know what is supposed to be wrong with it.
>>101795577
>/ldg/ is kind of a worse /DALL-3/ lately
An influx of nubs trying out the new thing, not knowing hot to fit in will do that.
>>
File: image.png (69 KB, 799x484)
69 KB
69 KB PNG
Midjourney CEO seething
>>
>>101795706
64 gpu at ONCE?!!!!!!!!

:o
>>
File: Real_Boris_Vallejo_Work.jpg (228 KB, 743x1037)
228 KB
228 KB JPG
>>101795120
That doesn't look like a Boris Valejo work at all though? Guess that flux doesn't know that style either
>>
>>101795691
nice
>>
File: CatJak_00153_.png (2 MB, 832x1280)
2 MB
2 MB PNG
>>101795704
Leg, nipples and he's not supposed to be furry, the mask has to be metal
These are done without negative guidance node
>>
>>101795691
Flux seems to do beautiful elves really well.
>>
>>101795728
>beautiful
>>
>>101795716
I really wish people who can train LoRAs would have eyes for more artistic eyes rather than just anime and vidya
>>
>>101795722
You will never get nipples in Flux that look normal. You need to run a nipple mask from Pony on top of them.
>>
>>101795737
need Dafne Keen pony lora
>>
>>101795730
I agree, she should be black.
>>
File: 1695082459867454.jpg (145 KB, 1074x663)
145 KB
145 KB JPG
>>101795706
Idogram nigga is also cooking + seething
>>
>>101795737
Could say the same for the people training models. Pony looks like fried slop, auraflow is slop, flux is slop. PLEASE TRAIN ON MORE SYNTHETIC DATA DADDY AGHHHHHHHHHHHH
>>
>>101795738
>You will never get nipples in Flux that look normal. You need to run a nipple mask from Pony on top of them.
I am pretty sure we just need to find the baked in positive prompt that erases them and counter it.
>>
>catjak still ahead of the pack
>>
they told me the schnell model couldnt do text
>>
>>101795750
Go woke, get broke. Fact is.
>>
>>101795750
kek
>>
>>101795746
Yes please!
>>
File: FD_00303_.png (1.41 MB, 1024x1024)
1.41 MB
1.41 MB PNG
>>101795756
I don't think that actually exists though. I think they ran an AI pass over all the nipples to make them look like bee stings before they trained the set.
>>
>>101795750
they are done, especially with the flux bias removal, now our local model can make any styles it knows pretty easily >>101793059
>>
File: 00013-835718714.png (2.85 MB, 1280x1920)
2.85 MB
2.85 MB PNG
>>
File: ComfyUI_02660_.jpg (912 KB, 1792x2304)
912 KB
912 KB JPG
>>
>>101793059
Awesome
>>
>>101795706
pseudoterminal is a huge sperg desu
>>
>>101795775
100% sure there is training in there that shouldn't be there :^)
>>
>>101795777
Only service worth using at this point is NAI for anime (and even than you're shit out of luck if the style/characters you like weren't in the data, and pony is good enough if you know what you're doing) Flux is competitive with midjourney/dall-e/etc, no surprise they are seething.
>>
>>101795794
very nice
>>
Keep it rolling...
>>101795805
>>101795805
>>101795805
>>
>>101795794
What happens with "DOF" in negative?

>>101795777
checked
>>
>>101795706
>>101795750
I'm glad people are losing money but it's bad other people are getting money, Fuck this.
>>
>>101795706
>>101795750
When they finally think they are going to get that imggen oligarchy going, flux barrels right through and ruin their plans. Now they have to keep pushing forward, or they will eat dust. Competition is good for everyone, and this seething is the proof.
>>
Why didn't you guys finish your previous thread
>>
File: ComfyUI_10522_.png (2.19 MB, 1440x1120)
2.19 MB
2.19 MB PNG
>>
>>101795827
sdg schizo baking early, anons refuse to ignore his thread
>>
>>101795841
just discord fag things
>>
>>101795827
They are finished, the proper way
>>
>>101795577
>my gens were never up to snuff with the quality you guys make
That's very sweet of you, I appreciate it.
>>
>>101796044
The collages are very good too. I'm very drunk now though and going to bed.
>>
>>101794987
>https://www.reddit.com/r/StableDiffusion/comments/1emy7uv/comment/lh7e9zl/?utm_source=share&utm_medium=web2x&context=3

https://github.com/asagi4/ComfyUI-Adaptive-Guidance
>>
>>101796163
thanks anon, have you tried it? does it work with flux?
>>
>>101796215
Yeah it works, here's the workflow: https://civitai.com/models/625042?modelVersionId=706397
I haven't messed with the workflow or compared with the other, but as it says there, it increases the gen time by about 3x vs. normal.
>>
If you missed it...
>>101795805
>>101795805
>>101795805



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.