[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: tmp.jpg (933 KB, 3264x3264)
933 KB
933 KB JPG
Discussion of free and open source text-to-image models

Previous /ldg/ bread : >>101792978

>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
>>
nice collage
>>
File: delux_nf_00022_.jpg (703 KB, 1344x960)
703 KB
703 KB JPG
>mfw
>>
File: FD_00003_.png (1.5 MB, 1024x1024)
1.5 MB
1.5 MB PNG
>>101795805
shocked we got an image local model before we got PromisedChan

picrel is an nvidia themed birth
>>
File: IMG_00082_.png (677 KB, 1024x768)
677 KB
677 KB PNG
>>
>>101795828
replace the kid through a GPU lmao
>>
File: FD_00273_.png (2.04 MB, 1024x1024)
2.04 MB
2.04 MB PNG
>>101795805
Sugoi!
>>101795716
How's this?
>>
File: ComfyUI_00399_.png (2.44 MB, 1536x1024)
2.44 MB
2.44 MB PNG
>>
File: ComfyUI_01881_.png (1.93 MB, 1024x1024)
1.93 MB
1.93 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

And here's some examples:
>A 1700s painting of a portrait of Hatsune Miku
https://imgsli.com/Mjg1Nzk5

>Hatsune Miku, 80's anime drawing style
https://imgsli.com/Mjg1Nzg2

>A watercolor ink pen outline painting of Hatsune Miku by Dean Crouser
https://imgsli.com/Mjg1ODI5

>a Boris Valejo fantasy painting of Hatsune Miku dressed as sailor moon
https://imgsli.com/Mjg1ODc4
>>
File: 00320-2024-05-06p.jpg (3.33 MB, 3072x3072)
3.33 MB
3.33 MB JPG
>>101795821
It's fun being rent free in this loser sperg's head
How is he still so shit at this when he put in 3x the time I put into flux
>>
File: IMG_00042_.png (1.57 MB, 1920x1080)
1.57 MB
1.57 MB PNG
>>
Why does debo keep trolling
>>
>>101795861
super super cool
>>
File: ComfyUI_10541_.png (2.61 MB, 1600x1200)
2.61 MB
2.61 MB PNG
>>101795848
alright we get it, have an upvote good sir1
>>
File: ComfyUI_00051_.png (934 KB, 832x1216)
934 KB
934 KB PNG
halloween is coming
>>
>>
File: FD_00311_.png (1.16 MB, 1024x1024)
1.16 MB
1.16 MB PNG
This one was weird to see in the preview. It genned a clearly anime hatsune miku then half way through switched to real.
Gives us a better understanding of how it's achieving this, I think
>>
>>101795805
>picked the best gens from last thread
Good choices
>>
File: f0016-jak.png (3.38 MB, 2048x2432)
3.38 MB
3.38 MB PNG
>>101795863
Nobody cares about him or the trash he makes and since he's not getting attention he's trying to shit up this thread
Might be worth putting the pastebin in the OP because this isn't a place that has been friendly to him from the very beginning.
>>
>>101795840
>How's this?
that looks better yeah
>>
>>101795858
the fact that a single fag posting 3 letters puts you in this much of a tizzy is fucking hilarious. serious feminine energy
>>
>>101795848
Someone should make a rentry for all the flux tricks anon has found
>>
File: ComfyUI_00418_.png (2.16 MB, 1536x1024)
2.16 MB
2.16 MB PNG
>>
>>101795848
Based
>>
>>101795888
Are you the baker
>>
>>101795906
I was a long time ago not anymore
>>
File: ComfyUI_00142_.png (1001 KB, 1216x1024)
1001 KB
1001 KB PNG
>>
>>
File: FD_00314_.png (1.65 MB, 1024x1024)
1.65 MB
1.65 MB PNG
>>101795917
Crab Maga
>>
>>101795848
what do i need to add to make negative prompts still work tho
>>
File: ComfyUI_00261_.png (1.25 MB, 1344x768)
1.25 MB
1.25 MB PNG
>>
>>101795931
I guess they don't work anymore at GuidanceNeg = 10 right? You've tried it?
>>
File: CatJak_00161_.png (2.01 MB, 832x1280)
2.01 MB
2.01 MB PNG
>>
>>101795848
I couldn't get your example to work. It seems to be completely ignoring the prompt and generating random images.
>>
File: ComfyUI_00052_.png (929 KB, 832x1216)
929 KB
929 KB PNG
>>
>>101795945
can you show a screen of your ComfyUi workflow so I can see what's goin on? And do you use schnell? Doesn't seem to work on that one, only dev does.
>>
File: ComfyUI_10547_.png (2.38 MB, 1440x1120)
2.38 MB
2.38 MB PNG
>>
File: screenie(53).jpg (1.05 MB, 2289x1114)
1.05 MB
1.05 MB JPG
>>101795945
It's not random. The outputs are consistent, even if it's now what you think they should be based on the prompt.
The thing is most of these are short and basic prompts. The more detail you add the closer it becomes like >>101795840
>>
File: ComfyUI_10548_.png (2.45 MB, 1440x1120)
2.45 MB
2.45 MB PNG
>>
File: ComfyUI_00053_.png (1.17 MB, 832x1216)
1.17 MB
1.17 MB PNG
>>
>>101795893
*I should have started compiling a rentry
>>
File: ComfyUI_10550_.png (2.5 MB, 1440x1120)
2.5 MB
2.5 MB PNG
>>
>>
>>101795805
Hooray for AI boobies!
>>
File: CatJak_00163_.png (2.1 MB, 832x1280)
2.1 MB
2.1 MB PNG
>>
File: COD648~1.png (1.37 MB, 1024x1024)
1.37 MB
1.37 MB PNG
>>101795848
why is the prompt a primitive and not a Clip Text Encode prompt?

trying to understand things plus i like to output the prompt to the filename
>>
File: FD_00321_.png (1.06 MB, 1024x1024)
1.06 MB
1.06 MB PNG
>>101795893
>>
>>101796053
>why is the prompt a primitive and not a Clip Text Encode prompt?
becuase the primitive is linked to the clip text encode prompt, so that you can input a text only once and it'll be activated on both t5xxl and clip_l
>>
>>101796053
So the t5xxl and the clip_l can be the same prompt and you don't have to enter it into 2 boxes
>>
File: ComfyUI_10552_.png (2.3 MB, 1440x1120)
2.3 MB
2.3 MB PNG
>>
hello /sdg/, excited for another day of avatarfagging and posting the same shit gens 10 times in a row?
>>
File: CatJak_00162.png (1.92 MB, 832x1280)
1.92 MB
1.92 MB PNG
I'm no longer using that negative encode trick
>>
>nigbo absolutely fuming
>>
File: ComfyUI_10554_.png (2.25 MB, 1440x1120)
2.25 MB
2.25 MB PNG
flux is so weird
>>
File: FD_00304_.png (1.34 MB, 1024x1024)
1.34 MB
1.34 MB PNG
>>101796080
Is it really avatar fagging if we are all fagging as Hatsune Miku and Donald Trump?
>>
Hatsunald Mump
>>
>>101796096
aww she's so cute with that blonde hair
>>
File: CatJak_00165_.png (2.02 MB, 1024x1024)
2.02 MB
2.02 MB PNG
>>101796096
It's a schizo that dedicated 2 years of his life to shitting up the general just laugh at him
>>
never forget what the /sdg/ refugees took away from us
>>
>>101795861
wow, is that the 0 bit model?
>>
File: FD_00326_.png (1.19 MB, 1024x1024)
1.19 MB
1.19 MB PNG
>>101796091
Nah he came to the light, he was wrong but now understands our future in Flux
>>101796113
I am aware of who it is I just saw the excuse to post my Trumpsune Miku and I took it.
>>
File: FD_00005_.png (802 KB, 1024x1024)
802 KB
802 KB PNG
>>
File: ComfyUI_10558_.png (2.24 MB, 1440x1120)
2.24 MB
2.24 MB PNG
>>
Flux can't gen people resting their feet and/or legs on tables. It just can't.
>>
File: file.png (732 KB, 542x703)
732 KB
732 KB PNG
CONCEPT ART ANON from last threadhere.
Sorry I was out for a bit.

>>101795370
Yes I did change the seed sorry

This is a proper comparison of default workflow vs the anti-slop workflow, same seed, same prompt (it's in the description.)
https://imgsli.com/Mjg1ODk1
>>
>>101796142
I am finding if I accidentally have unicode in the description, then I wind up with the grid pattern.

> Jasper Johns, The Critic Sees, 1961, Sculp-metal on plaster, glass, 3 1\u20448 \u00d7 6 1\u20442 \u00d7 2 1\u20448\u2033."
>>
>>
>>101796148
What happens with seed 667?
>>
File: FD_00328_.png (1.7 MB, 1024x1024)
1.7 MB
1.7 MB PNG
Let me tell ya folks, this is the greatest fusion in the history of America, maybe even the worldo, sugoi!
>>
>>101796159
>seed 667
kek nice
>>
>>101796159
I don't get it
>>
File: ComfyUI_10559_.png (2.26 MB, 1440x1120)
2.26 MB
2.26 MB PNG
>>
File: 1720829559326.png (1.12 MB, 1024x1024)
1.12 MB
1.12 MB PNG
>>101796159
It outputs this??
>>
https://github.com/asagi4/ComfyUI-Adaptive-Guidance
Looks like "Adaptive Guidance" is like CFG but without the 2x speed decrease, the problem is probably that is DynamicThresholding won't work on that shit so it's useless at the moment.
>>
>>101796182
kek, I loved that auraflow fiasco arc too
>>
File: FD_00335_.png (1.46 MB, 1024x1024)
1.46 MB
1.46 MB PNG
>>101796164
I've peaked.
>>
File: FD_00001_.png (1.06 MB, 1024x1024)
1.06 MB
1.06 MB PNG
>>101796222
Checked. Time for some new ideas.
>>
File: file.png (2.64 MB, 832x1216)
2.64 MB
2.64 MB PNG
Any other ideas for proooompting???? I am now very high enerdgy after discovering that flux can indeed do things other than slop.
>>
File: FD_00336_.png (1.5 MB, 1024x1024)
1.5 MB
1.5 MB PNG
>>101796222
I was wrong
>>
It slays me it was trained on modern art.
>>
>>101796237
I haven't tried yet, does DOF as a negative keep everything in focus?
>>
>>101796237
maybe turn him into a gooner death squad member
>>
>>101796253
no unfortunately, the model doesn't seem to understand the blur, DOF concept
>>
File: ComfyUI_10567_.png (2.55 MB, 1600x1200)
2.55 MB
2.55 MB PNG
>>
>>101796230
Very pleasing palette
>>
File: file.png (2.47 MB, 832x1216)
2.47 MB
2.47 MB PNG
>>101796253
Doesn't do anything
>>
File: FD_00308_.png (876 KB, 768x768)
876 KB
876 KB PNG
>>101796230
>Time for some new ideas
No
>>
>>101796253
Leave it alone dude, we have told you a billion times you can't do anything about it. This is just how flux is
>>
File: CatJak_00170.png (2.17 MB, 832x1280)
2.17 MB
2.17 MB PNG
Good night
>>
>>101796314
The son of a bitch did it
>>
>>101796313
blurry images get captioned, its there, just that negative prompt workaround isn't that good
>>
>>
>>101796313
We shall see :^)
>>
>>
>>101796348
Yup, and when you see people posting consistent images with no dof you will know it's been cracked. Until then stop asking. It's been a week.
>>
File: ComfyUI-Flux_00002_.png (1.71 MB, 1024x1024)
1.71 MB
1.71 MB PNG
>loading in lowvram mode
>loading in lowvram mode
>loading in lowvram mode
>>
Why are some anons so good at cracking models?
>>
>>101796364
Determination, grit, massive cocks among other things
>>
File: file.png (2.54 MB, 832x1216)
2.54 MB
2.54 MB PNG
Verification not required
>>
File: FD_00349_.png (1.87 MB, 1024x1024)
1.87 MB
1.87 MB PNG
>>101796364
>>
I'm still in awe over how talented and consistent some anons have been
>>
a second pass with an older generation model does not make sense considering none have the same level of detail, presumably because flux is 16ch
>>
>>101796428
XL does but the gen looks worse if you do a full pass. If you mask bits, like the nipples, you can glue them back on with Pony while keeping the superior Flux gen.
>>
File: 1700115362598684.jpg (76 KB, 1292x738)
76 KB
76 KB JPG
What the flux ?
>>
>>101796237
Give proompt for this one
>>
>>101796469
Lmao, nice gen
>>
MODS
>>
>>101796469
definitely not nips theyre not even the right color
youre safe
>>
File: ComfyUI_00950_.png (1.8 MB, 832x1216)
1.8 MB
1.8 MB PNG
>>101796478
>ultra Wide shot, AAA artstation photobash concept art of Japan tokyo metropolitan police tactical unit with cyberpunk aesthetic responding to a shooting with rifle raised and at high alert and high ready tactical posture, science fiction with exoskeleton suit and tactical equipment, military realism, highly detailed, downtown tokyo, fast ballistic helmet with armor panel scorings. The photo of the officer is with great intensity with motion blur. RAC, high cut ballistic helmet, plate carrier, backpack with SATCOM radio inside. The officer is maneuvering and the whole image is highly dynamic.

Sorry my proooompts are long as fuck
>>
File: Image.png (534 KB, 896x576)
534 KB
534 KB PNG
>>
>>101796479
you sick fuck...
>>
just report it
>>
File: FD_00004_.png (1.74 MB, 1024x1024)
1.74 MB
1.74 MB PNG
>>
File: ComfyUI_00767_.png (877 KB, 1344x768)
877 KB
877 KB PNG
>>101796469
>post similar gen several days ago.
>Get two (You)s
>Someone asks for prompt
>Provide it
>several days later someone posts similar promot
>Gets four (You)'s including mine

I don't care if it's petty. Those (You)s belong to me.
>>
>>101796237
>>101796126
>>101795917
>>101795848
>>101796314
Nice
>>
File: ComfyUI_00700_.png (861 KB, 1344x768)
861 KB
861 KB PNG
>>
File: ComfyUI_00097_.png (851 KB, 1152x896)
851 KB
851 KB PNG
>>101796559
in that case I'll claim credit for being the first to post anime screenshots like this, so now all the (You)s are mine, get fucked
>>
>>101796576
Well I invented anime so all the (You)s belong to me.
>>
File: 1.png (777 KB, 1920x953)
777 KB
777 KB PNG
>>101795952
>>
>>101796579
I am God and created the universe. There is no (You) that doesn't belong to me.
>>
File: ComfyUI_00095_.jpg (115 KB, 1024x1024)
115 KB
115 KB JPG
I was told Flux was DALL-E 3 level at foot stuff.
DALL-E 3 never fucked up to this degree.
>>
File: 1703595551076341.jpg (69 KB, 768x1024)
69 KB
69 KB JPG
>>
>>101796599
>I was told Flux was DALL-E 3 level at foot stuff.
who? who told you that? there's atleast 10 footfags a week that say flux is bad at feet.
>>
>>101796559
Thank you sir for proomt here's your (you).
I didnt changed a thing in proompt
>>
>>101796599
That's amazing.
>>
>>101796602
>This image is a high quality 2011 screenshot from a TV anime which is visually striking for its rich pastel color pallet soft line art and three tone shading creating a soft and inviting visual design. The image shows the two main characters in a locker room, one of the girls is pulling up a pair of tiny pink cute panties over her wide hips and plump bottom causing them to ride up her butt crack slightly, the other is pulling her shirt over her head, undressing, The view of the scene is partially obscured by black lines giving the impression that the camera is behind a grill or vent. The subtitles read, "I think I heard something from that locker."
>>
>>101796611
this fag https://desuarchive.org/g/thread/101787894/#q101788095
>>
I need several computers now.

One for Flux, one for other models, one for LLM, and of course one for gayming.
>>
>>101796142
Love this
>>
File: 1717540251226630.jpg (242 KB, 576x1024)
242 KB
242 KB JPG
Comic
>>
>>101796584
>steps 4 for flux dev
that's too low, it should be at least at 20
>>
>>101796655
Why can't you just have one PC with lots of storage and multiple GPUs? It would unironically be cheaper.
>>
>>101796655
why not just have one for ai and one for gaming?
>>
>101796660
kill yourself
>>
>>101796666
BE GONE, SATAN, TAKE THIS CHECK AND GO
>>
>>101795939
nigga can i get the prompt for this?
>>
File: ComfyUI_01129_.png (900 KB, 824x1224)
900 KB
900 KB PNG
Everything else fucked up, but the ass is good.
>>
flux gave us story telling through prompt alone
>>
>>101796665
Yes, I know. I was testing Schnell previously. But, even if unfinished, you can tell that image looks nothing like Hatsune Miku.
>>
File: Capture.jpg (955 KB, 3473x1309)
955 KB
955 KB JPG
>>101795848
https://files.catbox.moe/3bdsif.jpg
Ok I did a XY plot between GuidanceNegative (X) and CFG (Y) and... you're not gonna believe this, but the default parameter I've chosen before (CFG 6 + GuidanceNegative 10) seems to be the sweet spot... I think I never got this lucky in my life what the hell?
>>
>>101796715
go for 20 steps and see if it still gives you something random
>>
>>101796733
Why is cfg 8 wider angle?
>>
File: 1710322827929507.jpg (25 KB, 473x649)
25 KB
25 KB JPG
>Photographer wait for perfect shot for 100s hours
You much time you need to make it anon ?
>>
>>101796750
there's always some point in the CFG scale where you start to get something completely different compared to before, when that happens it's a sign we shouldn't go further I guess
>>
File: ComfyUI_00013_.png (1.22 MB, 1024x1024)
1.22 MB
1.22 MB PNG
just got back holy shit the differences in these renders with these settings changes
this is an old version of prompt
>>
Where are these negative guidance nodes coming from?
>>
File: ComfyUI-Flux_00003_.png (1.7 MB, 1024x1024)
1.7 MB
1.7 MB PNG
>>101796769
and newer config version like i actually had asked for tekken originally (and that first one was closest I got)
>>
File: file.png (2.55 MB, 832x1216)
2.55 MB
2.55 MB PNG
aaaaaaaaaaaaaa IM PROOOMPTING
>>
>>101796773
>Where are these negative guidance nodes coming from?
Comfy added them in? I'm not sure I understand the question kek
>>
>>101796773
Some anons are not using it like
>>101796084
>>
>>101796769
>>101796783
prompt "a gorgeous blonde woman wearing a black bodysuit playing a tekken cabinet in a busy futuristic arcade full of arcade cabinets and neon signs"
>>
>>101796769
>>101796783
>these settings changes
what settings changes exactly?
>>
why does a prompt of just "candid" generate kids 100% of the time? it even made a topless girl
>>
>>101796787
Oh so I just update comfy ui?
>>
>>101796799
yeah I guess
>>
>>101796660
thats a strange creature
>>
>>101796796
>>101796799
Dynamic Threshholding (its in comfyui manager) and changing the workflow in comfyui
heres how the guy on here (whos a redditfag) did it
https://www.reddit.com/r/StableDiffusion/comments/1enm9og/comment/lh79ucw/
>>
>>101796733
Can you share the workflow for xy plot? I don't know how to into it on Comfy
>>
>>
File: FD_00378_.png (1.57 MB, 1024x1024)
1.57 MB
1.57 MB PNG
>>
>>101796816
>>101796799
>>101796796
>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

save that png then drop it into comfyui with comfyui manager installed
>>
File: FLUX_00066_.png (1.34 MB, 1024x1024)
1.34 MB
1.34 MB PNG
>>
>>101796838
Why are you spamming this shit and ban evading?
>>
File: ComfyUI_00193_.png (1.05 MB, 1024x1024)
1.05 MB
1.05 MB PNG
>>101796756
Flux doesn't know how to generate an upside down bird, nor does it know what a kingfisher's head looks like, but otherwise not a bad effort
>>
>>101796846
He just wants attention. Don't give it to him.
>>
File: Capture.jpg (111 KB, 3177x618)
111 KB
111 KB JPG
>>101796820
Sure, here it is: https://files.catbox.moe/yse2v7.png
You need to install the "ComfyUI-nodes-hnmr" node on manager to be able to use the XYZ plot though
>>
>>101796783
Did you keep the CFG 6 + GuidanceNegative 10 values or did you change it a bit? seems like it's the sweet spot >>101796733
>>
File: ComfyUI_00116_.png (1.48 MB, 1024x1024)
1.48 MB
1.48 MB PNG
>>101796740
>>
>>101796399
>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.
No, the 2x is for CFG + DyanmicThresholding, and the 3x is for PrepNegGuider, for the Adaptive Efficiency it's supposed to be a 1x speed decrease, I'll try it out
>>
>>101796880
I think you should update comfyUi, something's wrong but I don't know what
>>
>1x speed decrease
>>
>>101795871
Catbox plz
>>
File: ComfyUI_00066_.png (1.2 MB, 1216x832)
1.2 MB
1.2 MB PNG
>>
File: 1707159653237000.jpg (225 KB, 1536x1536)
225 KB
225 KB JPG
>>101796847
Nice mine was not getting still water right .
Picrel is from google imagen
>>
File: ComfyUI_00180_.png (907 KB, 1024x575)
907 KB
907 KB PNG
>>101796864
Using your workflow I found that CFG 6 caused this weird grid effect, but dropping it down to ~3 gets rid of it (though obviously you lose some of the prompt adhesion benefits).
The grid effect only shows up on plain, untextured areas so obviously better prompting is a solution too, but it'd be nice to understand why it happens and try and figure a proper fix.
>>
>>101796907
is it really a vae decoder problem?
>>
File: ComfyUI-Flux_00004_.png (1.64 MB, 1024x1024)
1.64 MB
1.64 MB PNG
>>101796864
3.5 guidance+ / 10guidance- for that arcade one
heres another from that batch
>>
>>101796907
You weren't supposed to see the grid.
>>
>>101796957
>3.5 guidance+ / 10guidance-
and the CFG? CFG isn't guidance+
>>
File: file.png (75 KB, 552x557)
75 KB
75 KB PNG
>>101796961
uh idunno
>>
>>101796980
you can see the cfg on the Ksampler
>>
File: file.png (16 KB, 364x365)
16 KB
16 KB PNG
>>101796996
>>
>>101797001
ok thanks o/
>>
>>101796854
Thanks Anon,.
>>
>>
File: ComfyUI-Flux_00001_.png (1.96 MB, 1024x1024)
1.96 MB
1.96 MB PNG
>>101797009
no problem. never said i understood it all just using that other guys config ^_^
compare this on the new config to the next one is supposed to be a boris valejo of fairy on a mushroom dancing
>>
Am I "supposed" to be able to use an float8 model on my 3060 gpu? I never see it get loaded to the gpu

Something (not sure what) /automatically/ gets "manually casted" to bf16 even though I launch with args to make the unet (transformer) and text encoder be float8
>>
File: ComfyUI_00019_.png (1.38 MB, 1024x1024)
1.38 MB
1.38 MB PNG
>>101797041
old one
>>
>>101796907
>>101796949
>>101796958
Grid artifacts is not a VAE problem. It's an artifact of patch embeddings, inherent to all diffusion transformers that are using them. Most DiTs undergo additional training to suppress them to indistinguishable levels, although you can never get rid of them completely and they can be detected with certain filters like laplace edge detection, even if perceptually invisible.
>>
File: FD_00392_.png (1.11 MB, 1024x1024)
1.11 MB
1.11 MB PNG
>>
can I use my old rx590 (8 GB) for anything? Maybe just text or vae encoding/decoding?

otherwise thinking I'll connect that card to my screens and let my nvidia gpu do SD. would gain maybe 1 GB of vram like that
>>
>>101797074
>can I use my old rx590 (8 GB) for anything?
home media server, that's about it
>>
File: 1696510534607751.jpg (95 KB, 800x1170)
95 KB
95 KB JPG
I made muttmerican man praise me
>>
File: FD_00396_.png (1.28 MB, 1024x1024)
1.28 MB
1.28 MB PNG
>>
>>101797074
people have been using flux on old cards. i'm still using 1080ti just be patient and keep an eye on temps
>>101797083
i'd assume he meant fluxxin
>>
>>101797083
yeah, thought so. I started out on SD with that card when SD was new actually, but was thinking that it could be used for something at least (faster than the CPU for matmul et c)
>>
File: ComfyUI_00071_.png (1.1 MB, 1024x1024)
1.1 MB
1.1 MB PNG
me
>>
>>101797098
>i'd assume he meant fluxxin
i assume so too, so i gave him the honest answer for his 5 year old 8GB AMD card
>>
File: FD_00355_.png (1.39 MB, 1024x1024)
1.39 MB
1.39 MB PNG
>>101797089
That burger looks dry as fuck Needs some whipped cream.
>>
>>101797098
alright, thanks. will experiment a bit with it. thinking I'll just put the transformer on the Nvidia gpu and make it work.
>>
>>101797116
It's old for sure but he could probably put the VAE on it.
>>
could be it
>>
Chnnyposter doing it again
>>
File: FD_00173_.png (1.29 MB, 1024x1024)
1.29 MB
1.29 MB PNG
>>101797089
I made his rival.
>>
File: 1703800523761905.jpg (99 KB, 800x1170)
99 KB
99 KB JPG
>>101797089
And muttmericunt too
>>
>>101797167
would compare with regular images thrown through a pixelize node. then you can also customize the size of the pixels
>>
>>101797196
kek, he would get $100 tops for it now
>>
File: FLUX_00075_.png (1.21 MB, 896x1152)
1.21 MB
1.21 MB PNG
>>
File: FD_00403_.png (848 KB, 1024x1024)
848 KB
848 KB PNG
Flux can censor itself.
>>
File: ComfyUI_00074_.png (1.07 MB, 1024x1024)
1.07 MB
1.07 MB PNG
>>
>>101797260
Oh god I don't have any Neart weapons
>>
File: Capture.jpg (396 KB, 2911x1442)
396 KB
396 KB JPG
>>101796882
Ok I finally got what Adaptive Efficiency means, it's basically the same thing as CFG, but when we're reaching the end of the inference, when the pictures doesn't change much anymore, it reverts back to CFG = 1 -> 2x speed increase at that moment. It's cool but... DynamicThresholding doesn't work well with cfg = 1 and the result is this whiteish picture, I wish I could find a way to deactivate DynamicThresholding when the cfg = 1, is there a node that can do conditions and shit?
>>
>>101797269
it's over
>>
File: ComfyUI_00074_edit.png (1.1 MB, 1024x1024)
1.1 MB
1.1 MB PNG
>>101797269
couldn't stand it so i fixed it
>>
File: FLUX_00077_.png (945 KB, 896x1152)
945 KB
945 KB PNG
>>
>>101797294
NTA but this auction at $56 closed without any bids here in Sweden. Another was sold for like $45 though.
>>
File: file.png (467 KB, 1178x681)
467 KB
467 KB PNG
>Attach AutoCFG node to Dynamic Thresholding
>no more speed loss
Huh? Is it really that simple? I remember it slightly helping boost the speed gen with SDXL but holy shit.
>>
>>101797098
>keep an eye on temps
have you tried undervolting?
>>
File: Capture.jpg (108 KB, 3196x608)
108 KB
108 KB JPG
>>101797281
I put the workflow for those interested, if we can fix this we'll get a nice speed improvement
https://files.catbox.moe/w5jc8e.png
You need to download the adaptive node to make it work
>>
File: ComfyUI_01165_.png (1.17 MB, 1344x768)
1.17 MB
1.17 MB PNG
>>
>>101797407
and the output is the exact same?
>>
>>101797407
doesn't work, I got outputs as if it was cfg = 1
>>
>almost a mainstream term now
"tell me you spend a lot of time on the chans without telling me... et c"
>>
>>101797281
Interestingly if I use this workflow it gets text and prompt right really well, but like you said it's all washed out like it's on a faded piece of canvas.
>>
>>101797481
it kind of is, though. I've been seeing it places I wouldn't really expect to, like chats of big streamers and in youtube videos where there are young characters from video games and stuff like that
>>
Do you get speed improvements on Linux for flux as well? Or just less base VRAM use?
>>
>>101797481
Opus unironically uses this word without being prompted.
>>
>>101797059
Censored during the gen or after the gen? Post catbox either way
>>
>>101797281
>I wish I could find a way to deactivate DynamicThresholding when the cfg = 1, is there a node that can do conditions and shit?
maybe you could use this node?
https://github.com/theUpsider/ComfyUI-Logic
>>
File: _test.jpg (1.55 MB, 1344x2304)
1.55 MB
1.55 MB JPG
>>101797431
>>101797472
Well, it certainly does change the result but I'm pretty confused right now. Top - no autocfg and cfg = 5, slow as fuck and seems cooked, middle - autocfg and cfg = 5, bottom = autocfg and cfg = 1. Both middle and bottom results took the same amount of time as without any cfg hacks at all.
Prompt:
>An illustration inspired by the works of Jean-Baptiste-Siméon Chardin. The scene depicts a 18th-century classroom with a beautiful mature female teacher at the center, guiding a group of attentive young students. The teacher is dressed in period-appropriate attire. The classroom is filled with wooden desks, chalkboards, and books, all rendered with Chardin's characteristic focus on realistic textures and warm, muted colors. The words 'How to prooompt' are clearly written on the chalkboard. There is a speech bubble coming from the teacher's mouth with the text 'Don't be a retarded nigger.' The lighting is soft and natural, creating an atmosphere of calm and scholarly dedication.
Negative:
>anime, real, photo, 3d, render, depth of field, dof, blur
Also using ModelSamplingFlux with 0.75 max and 0.4 base shift because it kinda seems to make results nicer, both artworks and realistic stuff (my biased opinion though). Positive flux guidance at 3.15, negative at 10.
>>
File: FD_00401_.png (1 MB, 1024x1024)
1 MB
1 MB PNG
>>101797533
During.
>>
>>101797634
>>101797533
>catbox
https://files.catbox.moe/a2rcnf.png
>>
File: aa.jpg (2.82 MB, 3357x5353)
2.82 MB
2.82 MB JPG
>>101797623
For anime pictures, the output is really close to cfg = 1 so I don't see much change, but there's multiple autoCFG stuff like the warp drive I should try that one

Top: CFG6
Middle: CFG6 + Auto
Down: CFG1
>>
>>101797035
very cool
>>
File: FLUX_00084_.png (1.14 MB, 896x1152)
1.14 MB
1.14 MB PNG
more raunchy pensioner flashers
>>
>IP adaptors and better controlnets confirmed to be on the way
whos excited?
>>
>>101797756
IP-Adapter a shit
>>
File: FD_00015_.png (772 KB, 768x1024)
772 KB
772 KB PNG
>>
File: ComfyUI_01187_.png (1.11 MB, 1344x768)
1.11 MB
1.11 MB PNG
>>
File: file.png (3.09 MB, 1344x1534)
3.09 MB
3.09 MB PNG
>>101797623
Top - no cfg hacks at all so it's exactly the same as dynamic thresholding + autocfg + cfg = 1. Also took the same amount of time.
Bottom: dynamic thresholding + autocfg + cfg = 10. Still the same time. Need more testing with different styles.
>>101797714
The warp drive one didn't seem to work for me at all, even though it was my go-to node with sdxl. All the tests are done with the regular one, no idea about the other nodes
>>
>>101797756
Me
Myself
I
that's absolutely great
>>
I can't make it spell diarrhoea.
>>
>>101797811
good
>>
>>101797756
What is IP adaptors?
>>
>>101797042
>3060
I also got one of those. Thinking of buying another. Can you split the models up for them for FLUX?
>>
>>101797945
No. Best you can do is run the clip on one card and the model on another.
>>
>>101797949
Yeah, that's what I meant by splitting them to different GPUS. Sounds like a huge gain actually.
>>
File: fac011.jpg (416 KB, 1024x1024)
416 KB
416 KB JPG
>>
File: Capture.jpg (55 KB, 2801x180)
55 KB
55 KB JPG
>>101797789
Dude try this workflow: https://files.catbox.moe/8xt2m8.png

And download that script (it's a modified script from DynamicThreshold)
https://files.catbox.moe/eh22du.zip

You paste that shit into ComfyUI\custom_nodes\sd-dynamic-thresholding

What the script does now is to stop DynamicThresholding from working when AdaptiveGuider goes back to cfg = 1, for my example it goes back to cfg = 1 at 14/20 steps, it makes shit faster (50 sec instead of 1.05mn) and the image is better because it has seen some "normal" cfg for 1/4 of the time
>>
>>101797973
literally me
>>
File: FLUX_00094_.png (995 KB, 896x1152)
995 KB
995 KB PNG
>>
File: Flux_00076_.png (1.42 MB, 1024x1024)
1.42 MB
1.42 MB PNG
>>101793386
>>101792935
>>
>>101795805
How's the flux VRAM requirements now? Loadable with 8gb VRAM and 16gb RAM?
>>
>>101798156
OK but open her eyes
>>
need Dafne Keen pony lora
>>
>uses Linux, AMD, ROCm
Is it worth to pick up a used MI100? I literally see no one use any of the Radeon instinct cards
>>
keen deez nuts with your mouth
>>
>>101798190
turn your head and lift your sack
>>
File: Flux_00133_.png (975 KB, 1024x1024)
975 KB
975 KB PNG
>>101798156
yeah its hard to get crossed eyes wtf
>>
>>101796314
>>101794717

leaving without a chadbox? damn anon
>>
File: ComfyUI_00885_.png (1.21 MB, 1024x1024)
1.21 MB
1.21 MB PNG
>>101795805
>>
File: glif-flux-pro-prncs1.jpg (360 KB, 1024x1024)
360 KB
360 KB JPG
>>
>>
>>
Struggling a little to prompt modest clothing
>>
>>101798354
^ This was an attempt to prompt for pic related.
>>
>>
File: 1712073901642229.png (1.02 MB, 1499x2593)
1.02 MB
1.02 MB PNG
Sirs can flux do this i tried :(
>>
>>101798413
Can it do what?
>>
>>101798413
No.
>>
File: 1711276635909462.jpg (108 KB, 738x1292)
108 KB
108 KB JPG
>>101798422
It
>>
File: ComfyUI_00822_.png (1.24 MB, 1024x1024)
1.24 MB
1.24 MB PNG
>>101798291
imagine generating 55 pictures for anonymous posters on the internet instead of figuring out why windows isnt booting and the one you made before that was better anyway
>>
File: ComfyUI_00857_.png (1.16 MB, 1024x1024)
1.16 MB
1.16 MB PNG
>>101798475
and imagine you posted the wrong one on top of that!
>>
File: FD_00407_.png (1.16 MB, 1024x1024)
1.16 MB
1.16 MB PNG
>>101798413
>>
File: file.jpg (16 KB, 220x247)
16 KB
16 KB JPG
>>101798543
how horrifying
>>
>>101798556
One for your anus and the other for your urethra
>>
>>101798565
I had deflated but you made me turgid again.
>>
File: a.jpg (1.64 MB, 2201x4096)
1.64 MB
1.64 MB JPG
>>
File: dallegao (2).png (3 MB, 1024x1024)
3 MB
3 MB PNG
>>101798221
meanwhile, dalle: hey boss
>>
File: FD_00409_.png (1.46 MB, 1024x1024)
1.46 MB
1.46 MB PNG
>>101798581
>21 identical images
neat
>>101798413
I don't know what you want it to do but it makes some sexy devil girls
>>
>>101798596
That's not ahego either
>>
I made a tutorial to improve the inference speed by 25% if you are using CFG > 1 + DynamicThresholding
https://new.reddit.com/r/StableDiffusion/comments/1enxcek/improve_the_inference_speed_by_25_at_cfg_1_for/
>>
>>101797089
Skin too white for ogros americano
>>
I'm growing bored of flux. I need simple LoRA training support on a single 24gb GPU and a fine tune that has the nipples back.
>>
>>101798667
I'm not downloading and running random code from an anon, anon
>>
File: ComfyUI_00205_.png (1.73 MB, 1024x2048)
1.73 MB
1.73 MB PNG
>>101798413
Not too bad
>>
>>101798802
Ready? Set. Go!
https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_flux.md
>>
>>101798940
someone should make a PR to make DynamicThresholding inactive when the cfg = 1 desu, that was the point of this script
>>
>>101798667
>\r\n
>>
File: 1706390372727028.png (1.17 MB, 1024x1024)
1.17 MB
1.17 MB PNG
"in-game screenshot of the game World of Warcraft main menu" works, plus describing buttons or whatever.
>>
>>101795871
I thought flux couldn't make nice feet, this is nice.
>>
>>101798963
Where is there anything about 24gb on here?
>>
>>101795878
where are the bobs
>>
>>101799091
We've known this since it came out.
>>
>>101799142
Fuck you warcraft is my city
>>
File: ComfyUI_01205_.png (1.36 MB, 1344x768)
1.36 MB
1.36 MB PNG
>>101799150
>>
File: file.png (1.23 MB, 1024x1024)
1.23 MB
1.23 MB PNG
>>101799110
flux can do feet relatively well, just has issues with soles and sometimes fingertoes
>>
>>101799131
>https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_flux.md
>Where is there anything about 24gb on here?
There isn't because it's unnecessary. The docs aren't there to make your specific use case feel desired. Quantize if you OOM. diffusers is built by huggingface and essentially the standard
>>
File: 00155-AYAKON_124021214.png (2.98 MB, 1536x2560)
2.98 MB
2.98 MB PNG
>>
File: Sigma_12509_.png (2.67 MB, 2048x2048)
2.67 MB
2.67 MB PNG
>>
File: ComfyUI_01111_.png (964 KB, 1024x1024)
964 KB
964 KB PNG
>>
>>101799360
Creepy that I wouldn't know this was an AI image if I wasn't in the AI thread.
>>
File: ComfyUI_01214_.png (1.59 MB, 1344x768)
1.59 MB
1.59 MB PNG
>>
>>101799374
you know you can just add random captions to any image normally right?
>>
Ready to roll...
>>101799465
>>101799465
>>101799465
>>
>>101799472
poor anons been so inundated with fake ai images that they forgot photoshop is still a thing
>>
>>101799029
old reddit best reddit
>>
>>101798691
kek
>>
images got removed again
>>
>>101800308
Considering what they posted, seems about right
>>
bump



[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.