[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: 2024-08-06_00009_.png (2.41 MB, 1280x1024)
2.41 MB
2.41 MB PNG
Discussion of free and open source text-to-image models

Previous /ldg/ bread : >>101738379

>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
>>
Blessed thread of no collage frenship
>>
File: ComfyUI_30831_.png (524 KB, 864x1280)
524 KB
524 KB PNG
4-koma is really hard to prompt.
>>
File: 2024-08-06_00015_.png (2.39 MB, 1280x1024)
2.39 MB
2.39 MB PNG
>>101741569
lol thank you baker

also munch miku
>>
File: FD_00666_.png (1.15 MB, 1024x1024)
1.15 MB
1.15 MB PNG
>>101741591
Sometimes one gen is so good it needs to take centre stage.
>>
File: FD_00006_.png (1.59 MB, 1024x1024)
1.59 MB
1.59 MB PNG
>>
>>101741618
>>101741569
>Tfw Miku has moved on from music and is now part of the art movements
>>
File: Capture.jpg (232 KB, 2023x1615)
232 KB
232 KB JPG
Is there a way to put multiple guiders on SampleCustomAdvanced? I'd like to add the regular cfg on it aswell
>>
File: ComfyUI_01211_.png (1.51 MB, 1024x1024)
1.51 MB
1.51 MB PNG
>>101740870
>>101741104
so how can I change the CFG back to 1?
>>
File: Flux_00288_.png (709 KB, 728x728)
709 KB
709 KB PNG
Now we're getting somewhere
>>
File: ComfyUI_30832_.png (860 KB, 768x1536)
860 KB
860 KB PNG
I give up.
>>
File: ComfyUI_01249_.png (1.44 MB, 1024x1024)
1.44 MB
1.44 MB PNG
>>101741711
are you the Miku guy?
I got a question to your workflow, why is there some boxes purple?

I generated pic related with it and some boxes are purple in the workflow. wut it means?
>>
File: Capture.jpg (48 KB, 1090x711)
48 KB
48 KB JPG
>>101741796
yeah I'm the Miku guy from leddit kek, and you mean those purple box? it's just some deactivated shit when I was experimenting, you can make them active again by right clicking -> Bypass
>>
>>101741815
>you can make them active again
what does it change about the image? Imma ask first before bricking my shit again.
also any settings you can recommend, like steps, cfg etc. ?
like I dunno where to start there so would be interesting what the baseline is here if you know whatimsayin
>>
File: ComfyUI_Flux_00008_.png (1.6 MB, 960x1408)
1.6 MB
1.6 MB PNG
look what I found in the gamestop bargain bin
>>
>>101741794
I'm assuming you already tried describing what happens in panel 1,2,3 etc.
>>
Once we get functioning character LoRAs and contronets, it's over for he competition. We can probably expect a once in a blue moon good finetune too.
>>
File: ComfyUI_00307_.png (710 KB, 1024x1024)
710 KB
710 KB PNG
>>
File: ComfyUI_30829_.png (757 KB, 640x1280)
757 KB
757 KB PNG
>>101741931
Yup.
>>
File: Untitled.png (220 KB, 1154x374)
220 KB
220 KB PNG
>>101741931
>>
File: ComfyUI_00314_.png (865 KB, 1280x1280)
865 KB
865 KB PNG
>>101741896
>rated E
>>
File: ComfyUI_02378_.png (1.21 MB, 1024x1024)
1.21 MB
1.21 MB PNG
>>
There was an anon one or two threads back who plugged latent from SDXL to Flux, did anyone try that?
>>
>>101741887
>what does it change about the image? Imma ask first before bricking my shit again.
look at the box names, for example there's "Load Lora", it's useful if you want to add loras on flux but for the moment there's not much loras so it's useless

The other purple box is ModelSamplingFlux, it's a parameter that can make the model more creative, or you can use it to force it to adhere more to your prompt, you can activate it and try some values to see what it does

>also any settings you can recommend, like steps, cfg etc. ?
on a 1024x1024, 20 steps is fine, but like every other model, the more the merrier

for CFG I'm actually experimenting it, I can see that at cfg = 3 it makes miku black with dreadlocks, but the huge drawback is that it add a lot of saturation, even with the "hack", and I'm trying to find better parameters for the DynamicThresholding hack right now

Tbh, I think the best solution is using this workflow, it has PerpNegGuider, that's basically CFG but better because you can put cfg = 1 and you'll get better prompt adherance + you can use negative prompt, DynamicThresholding won't work on that one though because it was designed only for the "classic" CFG
https://civitai.com/models/625042/efficient-flux-w-negative-prompt
>>
File: cot.jpg (28 KB, 463x392)
28 KB
28 KB JPG
I'm gonna buy another 12 GB card
>>
>>101741956
? flux wasn't out at that time
>>
>>101741971
and tape them together?
>>
File: Flux_00300_.png (1.01 MB, 720x1504)
1.01 MB
1.01 MB PNG
>>101741945
Might be getting somewhere
>>
File: ComfyUI_00515_.png (1 MB, 1024x1024)
1 MB
1 MB PNG
>>
File: 1722567712507.png (1.03 MB, 1024x1024)
1.03 MB
1.03 MB PNG
>>
>>101741979
It was today or yesterday anon.
>>
>>101742001
That was eons ago
>>
File: ComfyUI_00334_.png (996 KB, 1024x1024)
996 KB
996 KB PNG
Literally 1984
>>
File: ComfyUI_00517_.png (841 KB, 1024x1024)
841 KB
841 KB PNG
>>
File: 2024-08-06_00012_.png (2.41 MB, 1280x1024)
2.41 MB
2.41 MB PNG
>>101741815
you are? thx for finding a way to make CFG work, makes some stuff way easier to achieve
>>
>>101741983
Yes

>>101742001
I read "threads" as "weeks" lmao
>>
>>101742030
lmao
>>
File: Thanks.jpg (11 KB, 197x255)
11 KB
11 KB JPG
>>101742032
you're welcome, desu I'm still not satisfied with the parameters on DynamicThresholdingFull, if I find something better I'll get back to y'all
>>
File: ComfyUI_02382_.jpg (518 KB, 1792x2304)
518 KB
518 KB JPG
>>
File: ComfyUI_00335_.png (900 KB, 1024x1024)
900 KB
900 KB PNG
I never mentioned Trump at all, I just prompted for an intimidating white man. What did flux mean by this?
>>
File: ComfyUI_01262_.png (1.3 MB, 1024x1024)
1.3 MB
1.3 MB PNG
>>101741966
>you can activate it and try some values to see what it does
hm doesnt really seem to improve the image.
>20 steps is fine
wasnt that also depending on what sampler you take like euler etc. ?
>https://civitai.com/models/625042/efficient-flux-w-negative-prompt
is that what you are using now?
I'll try it out.

also thank you for the explanation and help.
>>
File: 2024-08-06_00024_.png (2.49 MB, 1280x1024)
2.49 MB
2.49 MB PNG
>>101742045
I am using mimic_mode Constant instead of sawtooth, sawtooh made stuff wonky for me .. some anon in last thread said variability_measure STD instead of AD does less saturation (I still use AD) .. but overall the settings are fine I guess.
>>
File: 20240805T232948Z_00001_.jpg (490 KB, 1024x1024)
490 KB
490 KB JPG
>when you sleep well
>>
File: 40p2y19s3tgd1.jpg (1.94 MB, 9967x2393)
1.94 MB
1.94 MB JPG
>>101742079
>some anon in last thread said variability_measure STD instead of AD does less saturation (I still use AD)
It was me, you should look at the reddit post again, I updated the values, it has STD in it and other changes
https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/
>>
>>101742057
>War is peace
Get ready for ww3
>>
>>101742078
I use euler yeah, I don't see much difference with the others samplers (the one that are working kek)
>>
>>
You got a model that does text and basically anything you can imagine with good prompting and let me guess, you want more?
>>
>>101742118
It's amazing for photos and memes/jokes but it's lacking for art, Kolors is still a better art model if you don't need to generate text (which you usually don't)
>>
File: 2024-08-06_00026_.png (2.44 MB, 1280x1024)
2.44 MB
2.44 MB PNG
>>101742096
kek okay I see
>>
>>101742118
yes, I want giant stinky assholes in 4K HD and I want to prompt the exact shape and specification of her putrid shithole
>>
>>101741956
Not seeing why you would want to do that though? For what purpose?
>>
File: ComfyUI_01248_.png (1002 KB, 1024x1024)
1002 KB
1002 KB PNG
>>101742057
Kino
>>101742108
>(the one that are working kek)
which ones are working?
also how come they make no difference? in SD it usually changed the output image a lot depending on that sampler setting.
>>
>>101742145
>which ones are working?
try them by yourself, for example euler_cfg_pp makes the image completely black, some samplers just don't work with flux, I haven't tried them all though
>>
File: ComfyUI_02387_.jpg (593 KB, 1792x2304)
593 KB
593 KB JPG
>>
>>101742145
>also how come they make no difference? in SD it usually changed the output image a lot depending on that sampler setting.
probably because of the architecture, Flux is a DiT while SD models are unet, and desu that's a good thing, if you want to add something new like a flower pot, the whole picture won't change much, it'll just add the flower pot, that's cool
>>
File: Flux_00002_.png (842 KB, 1024x1024)
842 KB
842 KB PNG
>>101742153
also this other workflow you suggested works really good.
>https://civitai.com/models/625042/efficient-flux-w-negative-prompt

this is imo the gold standard for flux workflows.

>>101742171
interesting
>>
>>101742079
>but overall the settings are fine I guess.
for drawings yeah it works fine, but for realistic stuff you can see the oversaturation, that's something that needs to be fixed
>>
>>101742171
What does that mean for controlnets thought?
>>
>>101742181
For me it slows down my gens from 1s to 3s/it. Any idea why?
>>
Hello /long dick general/, is it necessary to use more steps with flux dev than with flux schnell? Schnell seems to require only 4 steps, but I'm not sure about dev.
>>
>>101742249
20 minimum for dev.
>>
flux runs fine on my 3090 but it's taught me that I valued aesthetics and vibes more than prompt adherence
I don't really care if the model gives me exactly what I specified, and I usually don't actually have an exact idea of what I want the composition to be anyway
I just want it to look nice and artsy and sovlful and I'm happy to let the model go in its own direction in pursuit of that rather than strictly obey me, in fact I prefer that
Seems like (with Flux and Aura and Lumina etc) things are moving in the opposite direction
Sad
>>
File: 2024-08-06_00031_.png (2.47 MB, 1280x1024)
2.47 MB
2.47 MB PNG
>>101742247
thats expected, if you raise CFG it will be about double the time to gen
>>101742249
yes dev wants 20+
>>
>>101742249
dev needs 20-50, schnell is a distilled/'turbo' model and dev is not
>>
File: Flux_00006_.png (1.69 MB, 1024x1024)
1.69 MB
1.69 MB PNG
>>101742247
dunno man I'm wondering the same.
what I noticed is that the workflows change how the GPU operates.

like the basic Flux workflow.
https://openart.ai/workflows/maitruclam/comfyui-workflow-for-flux-simple/iuRdGnfzmTbOOzONIiVV

seems to gen somewhat fast but it slows down everything else on the computer.
like I cant even watch a youtube video.
the Output is also somewhat worse.
what I also noticed with GPU Z is that it also doesnt create so much Heat and seems to draw less Power.
>>
File: FD_00099_.png (1.42 MB, 1024x1024)
1.42 MB
1.42 MB PNG
So who is the flux mascot? SD1.5 was Emma Watson. Flux is either Miku or Trump
>>
>>101742281
>20-50,
Is there any point going over 20?
20 looks pretty much "finished" to me.
>>
File: 3.png (2.31 MB, 920x1296)
2.31 MB
2.31 MB PNG
>>
File: 2024-08-06_00034_.png (2.46 MB, 1280x1024)
2.46 MB
2.46 MB PNG
>>101742315
def. Miku, Trump is only good for memes
>>
>>101742298
We desperately need a TensorRT workflow. It's horribly unoptimized without.
>>
>>101742247
>For me it slows down my gens from 1s to 3s/it. Any idea why?
>>101742298
>dunno man I'm wondering the same.
That's because PrepNegGuider is a new technique of CFG, it's also more expensive in terms of calculus
https://github.com/Perp-Neg/Perp-Neg-stablediffusion
>>101742278
>thats expected, if you raise CFG it will be about double the time to gen
no, it's also slow when at CFG = 1
>>
>>101742317
for some more complicated stuff you get errors without giving it a lot of steps
>>
>>101742337
oh ok
>>
so it really was just a skill issue, huh
>>
File: FD_00106_.png (1.14 MB, 1024x1024)
1.14 MB
1.14 MB PNG
>>101742321
You might be right
>>
>>101742361
for SAI
>>
File: ComfyUI_00543_.png (1.23 MB, 1024x1024)
1.23 MB
1.23 MB PNG
>>101742373
>>101742321
>>101742315
We have the tech no combine the two.
>>
>>101742422
delete this post
>>
File: ComfyUI_00072_.png (1.15 MB, 1024x1024)
1.15 MB
1.15 MB PNG
flux is pretty wild, very exciting to see things finally moving forward
are there any good guides out yet?
I'm curious about prompting, weights, all that jazz
(reply to this post or your mother will die in her sleep tonight)
>>
>>101742422
why am i horny and scared
>>
How does Flash Attention get used? Is it something I use with a distributed model, or is the model modified by Flash Attention by others?
>>
File: ComfyUI_00116_.png (1.07 MB, 1024x1024)
1.07 MB
1.07 MB PNG
>>101742437
no guides i know of but boomer prompting works well, as well as imitating the kind of descriptions AI comes up with when you ask them to describe an image
>>
File: Flux_00008_.png (1.31 MB, 1024x1024)
1.31 MB
1.31 MB PNG
>>101742437
just read the thread bro
>>
File: 39641682.png (2.4 MB, 1352x768)
2.4 MB
2.4 MB PNG
>>
File: 2024-08-06_00043_.png (1.37 MB, 1024x1024)
1.37 MB
1.37 MB PNG
>>101742437
this >>101742454
>(reply to this post or your mother will die in her sleep tonight)
ya my mother is dead already, so no luck on that

but really: guidance makes your prompt fly.. lower it and it will follow styles more, make it higher and it follows prompt logic better.. there are tricks to get cfg (and by that negative prompts) working but they are advanced (just scroll up abit)
>>
File: ComfyUI_00081_.png (1.24 MB, 1024x1024)
1.24 MB
1.24 MB PNG
>>101742453
>>101742454
haven't been to /g/ in a few months, since everything got stale after SDXL / CtrlNet
give me the damage report, are there two threads now? /ldg/ and /sdg/?

>no guides i know of but boomer prompting works well, as well as imitating the kind of descriptions AI comes up with when you ask them to describe an image
that seems to be the case, yeah
regarding specs, euler and 40+ steps seem to provide the best results, at least for realism
any other relevant tips I should be aware about?
>>
File: aaa.jpg (1.44 MB, 4040x2393)
1.44 MB
1.44 MB JPG
Ok guys, do not sleep on PrepNegGuider, this shit is really cool (if you are ok with the speed decrease), I provide you a workflow for those interested
https://files.catbox.moe/eriobu.png

Because we got the exact same output for regular CFG and PrompNeg CFG when we don't put any negative prompt, comfy should code this shit so that it revert back to regular CFG when we don't put negative prompt (for the speed increase)

https://perp-neg.github.io/
>>
File: FD_00113_.png (1.13 MB, 1024x1024)
1.13 MB
1.13 MB PNG
>>101742422
>>
>>101742494
based, thank you, sempai
>>
>>101742498
I used it to make anime girls non chibi, it works well.
>>
>>101742485
idgi, your example seems to show that it's pointless? the first image is fine already, the second one didn't improve upon it
>>
>>101742509
cars
>>
>>101742509
>idgi, your example seems to show that it's pointless? the first image is fine already, the second one didn't improve upon it
that's the point, it doesn't change the original output if you don't put any negative prompt, that's for the third picture you can see the difference
>>
File: 1803633435.png (2.27 MB, 872x1312)
2.27 MB
2.27 MB PNG
>>
File: ComfyUI_01088_.png (1.29 MB, 1024x1024)
1.29 MB
1.29 MB PNG
>>101742495
>haven't been to /g/ in a few months
heretic
>>
>>101742328
Is it not compatible with flux or something?
I wanted to use it with the PrepNegGuider posted here.
>>
>>101742498
in the third image the goat stopped being a billboard and just became a giant goat
>>
>>101742546
it was brutal
I still have flashbacks from the savage faggotry
all those avatarfags...
>>
File: Capture.jpg (33 KB, 912x494)
33 KB
33 KB JPG
>>101742569
true, you can change the strength of the negative prompt though, I think 1 is too high
>>
>>101742437
Didn't find any guides either besides going with boomer prompts like >> 101742453 said, I still need to dip more than my toes into flux (and upgrade my PC already)
>>
>>101742574
They're still jerking each other off in the other thread and haven't realized we all moved here.
>>
>>101742495
whats sdg?
>>
>>101742617
Sustainable development goals. Nothing else.
>>
>>101742601
which one do you decrease to just weaken it? cfg or neg_scale?
>>
File: ComfyUI_00490_.png (1.4 MB, 1024x1024)
1.4 MB
1.4 MB PNG
>>
>>101742641
don't touch cfg this shit fries the picture, it's the neg_scale that is responsible of the strength of the negative prompt
>>
>>101742610
yeah, same, haven't found anything yet

>>101742616
>>101742617
based
yuge /ldg/ W
>>
>>101742057
Why are you a murderer? Don't you know the law will hound you?
>>
File: FD_00115_.png (1.42 MB, 1024x1024)
1.42 MB
1.42 MB PNG
>>101742498
For that example couldn't you get a better result by simply prompting for an empty street?
For most examples, why can't you just describe the literal scene you want to see? Why do you NEED negatives?
>>
>>101742674
>Why do you NEED negatives?
negatives are important, if I want to remove a color, a cloth, even Miku's quills, you put that shit on negative prompts
>>
>>101742674
Not all concepts have opposites you can add like that.
I prompted for a figure, half of it was chibi, and I can't ask for "not chibi" because pink elephant, and realistic doesn't transfer to the figure but to the photo.
>>
>>
File: ComfyUI_00552_.png (874 KB, 1344x768)
874 KB
874 KB PNG
Trumps been looking kind of rough since Miku got the candidacy
>>
>>101742690
this .. the guy who made op post posted his workflow last thread, only made it possible with negatives
>>
>>101742674
text encoders (like t5xxl) are notoriously bad at negation, it's always better to put a word on a negative prompt rather than saying to the model "NOT this"
>>
>>101740951
Had a nice nap today, happy to share: https://files.catbox.moe/w2zjie.png

>>101742122
https://files.catbox.moe/15t9ir.png
>>
>>101742498
>get better gens by making my already slow gen time 3x worse
no pain no gain I guess
>>
>>101742742
it's twice as slow if you use regular CFG > 1 + DynamicThresholding or 3x slower if you use PrepNegGuider CFG = 1, desu I'd go for PerpNeg because it doesn't fry the picture at least
>>
File: 0.jpg (449 KB, 1024x1024)
449 KB
449 KB JPG
>>
File: ComfyUI_00492_.png (1.36 MB, 1024x1024)
1.36 MB
1.36 MB PNG
>>
>>101742674
>For that example couldn't you get a better result by simply prompting for an empty street?
Empty street means "No cars + No people", I just wanted "No cars"
>>
>>101742265
I'm like this too
The ideal situation for me would be being able to generate MidJourney starter images in a comfy workflow and then use SDXL/Pixart/Flux to do various modifications and upscales of the MJ image in an automated way

But I can't because MJ refuses to release a fucking API
>>
>>101742265
>>101742765
>The model gave me exactly what I asked for and I'm sad.
Go back to making SDXL slop then.
>>
File: 2024-08-06_00062_.png (2.81 MB, 1680x1056)
2.81 MB
2.81 MB PNG
>>
>>101742265
I am shocked that you put aura in there with the rest. It was a terrible turd that could not generate anything aesthetically pleasing.
>>
>>101742422
>Mump
Lmao
>>
>>101742713
Kek this is amazing, catbox?
>>
File: 4884.jpg (109 KB, 1200x858)
109 KB
109 KB JPG
>OSError: flux-schnell does not appear to have a file named config.json.
>Check out 'https://huggingface.co/flux-schnell/tree/main' for available files.

>no config.json in that tree
>>
>>101742858
https://files.catbox.moe/6lcf5s.png
>>
>>101742845
Oh no I don't like Aura either for the same reasons as you, I just meant that it's doing the same thing as Flux (attempting to have extreme prompt adherence somewhat at the expense of style using LLM dataset captioning)
they have the same goal, it's just that Flux actually achieved it while Aura didn't due to incompetence
>>
File: 1713085872442308.jpg (316 KB, 1024x1024)
316 KB
316 KB JPG
>>101742265
Finally someone else who agrees. Flux is overall compositionally more intelligent than Dall-E, but artistically so far behind. The negative effects of using synthetic ai-generated image data. soulless in - soulless out.
>>
>>101742869
>Aura didn't due to incompetence
All we've seen is a beta of a beta. isn't it just one guy?
>>
>>
File: kk.png (820 KB, 1024x1024)
820 KB
820 KB PNG
dawn lookin up
>>
File: ComfyUI_00497_.png (1.46 MB, 1024x1024)
1.46 MB
1.46 MB PNG
>>
>>101742862
post the full error
>>
File: [flux-dev]_00343_.png (1.16 MB, 1024x1024)
1.16 MB
1.16 MB PNG
>>101742722
Great job on the otter mage odyssey, I came back today and saw the rest of it.
Despite what the trolls in these threads say. FLUX's adherence to prompt (especially in terms of scene composition) is going to be the key to unlock the huge creative potential of AI images; A move away from the allure of "cool because it looks cool" towards new frontiers; like art with more "character" (for lack of better word)
>>
File: ComfyUI_00509_.png (1.48 MB, 1024x1024)
1.48 MB
1.48 MB PNG
>>
File: 2024-08-06_00066_.png (1.12 MB, 1280x1024)
1.12 MB
1.12 MB PNG
>>101742862
you know you need tu run it on comfy?
>>
>>101742862
Did you put the Schnell model in the unet or the checkpoints folder?
>>
File: 2024-08-06_00067_.png (1.09 MB, 1280x1024)
1.09 MB
1.09 MB PNG
>>101742925
it belongs in the unet folder, not the checkpoint folder
>>
>Flux doesn't use a Unet
>Put the Flux weights in the Unet folder

What did they mean by this?
>>
>>101742898
It's just the classic error stemming from HF's from_pretrained() naming reqs.
In this instance it's actually valid given that the file isn't provided
>>
File: WWWQ.png (1.3 MB, 1232x928)
1.3 MB
1.3 MB PNG
OM NOM NOM
>>
File: ComfyUI_30846_.png (3.83 MB, 1536x1536)
3.83 MB
3.83 MB PNG
>>
>>101742862
>>101742948
it's black-forest-labs/FLUX.1-schnell and you need to use subfolder=transformer unless you're loading the pipeline
>>
>>101742933
https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main

What unet dir?
>>
File: wewewe.png (1.38 MB, 928x1232)
1.38 MB
1.38 MB PNG
NOM NOM
>>
>>101742974
In the comfy UI folder. You put the weights in the Unet folder in YOUR comfy UI folder.
>>
>>101742498
>OverrideVAEDevice OverrideCLIPDevice nodes not found
what are these?
>>
File: ComfyUI_00510_.png (1.41 MB, 1024x1024)
1.41 MB
1.41 MB PNG
>>
>>101742992
I think you need to update comfy? and it's for putting the clip or VAE on another GPU if you have multiple GPUs
>>
File: FD_00398_.png (1.06 MB, 1024x1024)
1.06 MB
1.06 MB PNG
>>101742996
>>
File: ComfyUI_00567_.png (928 KB, 1344x768)
928 KB
928 KB PNG
>>
File: Flux_00036_.png (1018 KB, 1024x1024)
1018 KB
1018 KB PNG
>>101742498
>Positive prompt: A portrait of Hatsune Miku
>Negative prompt: cyan color, blue color, green color, pigtails, bangs, quilts, barrettes, rectangular hair clips
can't seem to be able to remove those fucking bangs on this bitch :(
>>
File: 2024-08-06_00073_.png (1.22 MB, 1280x1024)
1.22 MB
1.22 MB PNG
>>101742992
Those are custom nodes, you only need them if you use two GPUs, if you do, get ComfyUIManager and install the custom nodes..

alternatively use a more sensible default workflow like the ones provided by comfy
>https://comfyanonymous.github.io/ComfyUI_examples/flux/
>>
File: ComfyUI_00565_.png (1.03 MB, 1344x768)
1.03 MB
1.03 MB PNG
>>
>>101743042
kek, now that's fucking amazing, flux is really a blessing from the sky
>>
File: ComfyUI_30849_.png (3.81 MB, 1536x1536)
3.81 MB
3.81 MB PNG
>>
>>101742498
>comfy should code this shit so that it revert back to regular CFG when we don't put negative prompt (for the speed increase)
you can use Switch (Any) from https://github.com/ltdrdata/ComfyUI-Impact-Pack
It has condtional node that will do exactly as you want, too much of a brainlet to help you further though
>>
File: 2024-08-06_00075_.png (1.21 MB, 1280x1024)
1.21 MB
1.21 MB PNG
>>101743045
wth are you trying to achieve, hatsune miku is a strong prompt, its incredibly difficult to break even her style, what ya doing is putting a dodge viper motor into a toyota
>>
>>101742918
no, you don't. check their repos

>>101742970
thanks, will try that instead.
>>
>>101741715
change cfg value in ksampler node
>>
>>101743079
>hatsune miku is a strong prompt, its incredibly difficult to break even her style
it should be a way to remove the insane bias flux has on hatsune miku though
>>
File: 2024-08-06_00076_.png (1.36 MB, 1280x1024)
1.36 MB
1.36 MB PNG
>>101743097
it is, check last thread
>>
File: gghr.png (1.69 MB, 928x1232)
1.69 MB
1.69 MB PNG
>>
>>101742981
Not using Comfy
>>
>>101743133
Well, you don't really have a choice in this matter.
>>
>>101742884
the project got canned a day or two after flux's release
>>
>>101743148
Good. Piece of shit.
>>
File: ComfyUI_00573_.png (1.1 MB, 1024x1024)
1.1 MB
1.1 MB PNG
>>
>>101743143
How do you mean? It's just a HF-based script
>>
I'm new to Comfy and installed it just to run Flux.
I already did the my first image gen and it went alright.
The issue is that when I look in the terminal, there's the following message: "loading in lowvram mode 21989.075".

What does that mean?
I'm using a RTX 3090(24GB VRAM) and I have 64GB of DDR4 memory. Any ideas?
>>
File: Capture.jpg (337 KB, 3190x1388)
337 KB
337 KB JPG
>>101743045
PrepNeg still works way better at negatives than regular CFG + DynamicThresholding though
>>
So, what speeds for 1024x1024 gens are people getting on stuff on par with a 3080 for Flux? Feel like my iterations per second (or seconds per iteration) is slower than expected. Using the FP8 weights and no extra flags passed to comfy.
>>
File: ComfyUI_00574_.png (1.02 MB, 1024x1024)
1.02 MB
1.02 MB PNG
>>
>>101743083
https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux
optionally use pipe.enable_sequential_cpu_offload()
>>
https://github.com/pamparamm/sd-perturbed-attention
>Paper and demo suggest using CFG scale 4.0 with PAG scale 3.0 applied to U-Net's middle layer 0, but feel free to experiment.
Is this possible to use this PAG method on DiT architectures like flux?
>>
File: 2024-08-06_00089_.png (1.03 MB, 1280x1024)
1.03 MB
1.03 MB PNG
>>101743184
FLUX is so big that even 24GB VRAM is not enough to not go into lowvram mode, maybe if you turn on fp8 for both model and text encoder it will avoid that, but dont break your head about it, unless you have CUDA swapping turned on in your drivers the gen times will be fine even on lowvram mode
>>
File: ComfyUI_00578_.png (544 KB, 1024x1024)
544 KB
544 KB PNG
Flux only knows like Miku and Trump. Can't get megumin no matter how hard I prompt.
>>
File: ComfyUI_Flux_00042_.png (1.34 MB, 960x1408)
1.34 MB
1.34 MB PNG
>>101743184
you should be able to run the fp16 model with quants. did you quant at all?
>>
>>101743184
24GB of vram IS low vram.
>>
File: FD_00126_.png (1.29 MB, 1024x1024)
1.29 MB
1.29 MB PNG
>>
>>101743265
By quant do you mean FP8? Is there any difference between just loading it in FP8 using the FP16 version and using the pre quanted FP8 models out there?
>>
>>101743266
>>101743250
I see.
That's what they meant by THICC model.

>suddenly, I'm a vramlet
How could this happen, bros?
>>
File: ComfyUI_00072_.png (1.34 MB, 1216x832)
1.34 MB
1.34 MB PNG
>>101743254
That's because megumin a shit. That whole story a shit. The only one with a good aesthetic was darkness and she was a huge whore and should thus be crucified.
>>
File: FD_00128_.png (781 KB, 1024x1024)
781 KB
781 KB PNG
>>101743254
I don't know anime but I just described the image you posted. How's this?
>>
>>101743265
>quant

What is that?
>>
>>101743254
makes sense considering Miku is the oldest SAI project and its the same people
>>
File: 2024-08-06_00094_.png (1.03 MB, 1280x1024)
1.03 MB
1.03 MB PNG
>>101743254
>megumin
to obscure for the Kraut devs to care I guess, atleast it knows elderly Jackie Chan

>>101743305
>How could this happen, bros?
We all are now
>>
>>101743329
The one I posted was much closer to Megmin than yours.
>>
>>101743227
It was the subfolder option that was new to me. Didn't seem to have an effect though, even though it was accepted without protest.
>>
File: ComfyUI_00945_.png (1.28 MB, 1024x1024)
1.28 MB
1.28 MB PNG
>a 20 year old Donald Trump
Trump is so overcooked you can't change anything on him
>>
File: cfjhxfthj.jpg (15 KB, 620x200)
15 KB
15 KB JPG
>>101743297
there is an alternative model with fp8 precision instead of fp16. there shouldn't be too much of a difference vram wise but I haven't seen anyone compare them or anything

>>101743305
I have a 4090 and I am a vramlet kek

>>101743343
quantization. use less vram but takes more time is as layman as I can get for you. picrel where you change it
>>
File: 2ylwh87gzwgd1.jpg (170 KB, 1152x768)
170 KB
170 KB JPG
kek
>>
>>101743378
>I have a 4090 and I am a vramlet kek
A100 Bros rise up
>>
>>101743382
Actual reddit front page.
>>
>>101743364
yeah you only need that with like FluxTransformer2DModel. what you want is simply
from diffusers import FluxPipeline
pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16)
pipe.enable_sequential_cpu_offload()
>>
>>101743394
indeed
>>
File: 1.jpg (2.35 MB, 1904x3360)
2.35 MB
2.35 MB JPG
My irl friend thought this was a real image
>>
>>101743405
lies

you dont have an irl friend
>>
>>101743405
it is pretty convincing, did you edit this at all or is it just a raw gen?
>>
File: media_GUPmuEVXwAA2kb8.jpg (106 KB, 1024x1024)
106 KB
106 KB JPG
https://xcancel.com/Konan92_AI/status/1820518655450562588#m
Shot's fired
>>
>>101743401
Yes. That's what both their repos state.
>>
>>101743368
>can't change anything on him
>>101742422
>>
>>101743429
fucking based
>>
File: ComfyUI_00504_.png (1.39 MB, 1024x1024)
1.39 MB
1.39 MB PNG
>>
>>101743448
yeah you can change his hair, but not his face by making him younger, flux has way too much biases
>>
>>101742422
HO-LY-SHIT
>Mump
>>
>>101743429
Holy shit, what a fucking baby.
SAI are told exactly what they need to do to succeed and they still fuck it up everyone time. I don't understand how they can have the investment and resources they do and still not be able to produce anything resembling a good model. Even SDXL is complete shit save for the fact it was relatively easy to train.
>>
File: ComfyUI_00516_.png (1.38 MB, 1024x1024)
1.38 MB
1.38 MB PNG
>>
>>101743429
>Comparing a 12b API model (that's Pro you're using) to a 2b open weights model
Pretty sure SD3 has a 10b API model that's also massively dogshit.
>>
File: ComfyUI_00524_.png (1.17 MB, 1024x1024)
1.17 MB
1.17 MB PNG
>>
>>101743477
>SAI are told exactly what they need to do to succeed and they still fuck it up everyone time.
This, I don't know why they're salty about flux, they could do the same they have the architecture and the money to burn. I think they're just jealous of the flux team because they have way more balls than them

this is what Emad has to say about flux
https://xcancel.com/EMostaque/status/1819777262071226611#m
>fwiw I don’t think bfl team could have built/released Flux when at Stability (contrast aha) for… reasons
no shit, there's a reason the competent people left your cucked company
>>
>>
File: ComfyUI_00526_.png (1.15 MB, 1024x1024)
1.15 MB
1.15 MB PNG
>>
>>101743491
yeah, there's the SD3-8b we can try on the API and it's nowhere close to flux, he's just talking like a nigger faggot, as usual, the shitstorm that happened to him during the SD3M fiasco is completely deserved
>>
>>101743513
damn, looks like a trainer size.. boi needs to step up his game
>>
>>101743442
what's the problem then
>>
File: Flux_00307_.png (612 KB, 720x1504)
612 KB
612 KB PNG
What do the different weight_dtype do in comfy?
>>
>>101743541
i can tell you, if you do default its slow as dog balls.. if you do the fp8_blahblahblah then its much much faster
>>
>>101743506
I'd leave to if my open weights AI research focused company suddenly go coopted by middle management safety trannies and demanded I produce shit when I could be making quality progress.
>>
>>101743506
he tries so fucking hard to dodge responsibility, like it wasn't he who hired ella irwin and it wasn't he who dipped without dropping the model weights. he was the fucking ceo and he acts like he was just some outside observer with inside info. emad fucked over local so hard i hope he rots
>>
>>101743541
for flux that's how you switch to fp8 for half the vram usage and slight quality loss
>>
>>101743523
>>101743523
I don't have a twitter account. Someone tell him this.
>SD3 also has an API model and when we compare that to even Flux.1 Dev it's night and day. YOU need to do literally anything else, because making image generation models is not your field.
>>
>>101743458
How did you gen Sarah Jessica Parker? Every time I try I get a human woman.
>>
File: 6cagvjh62xgd1.jpg (329 KB, 2223x894)
329 KB
329 KB JPG
https://github.com/caleboleary/ComfyUI-Arc2Face
https://arc2face.github.io/
Could this work on flux? We wouldn't need loras anymore, just a picture of someone would be enough
>>
>>101743590
hot anthropomorphic camel woman

or something like that
>>
>>101743598
Loras are used for a lot more than just faces. For example, giant inverted milky nipples and specific artist styles.
>>
File: ComfyUI_00542_.png (1.4 MB, 1024x1024)
1.4 MB
1.4 MB PNG
>>
>>101743616
yeah true I oversimplified my sentence, I meant that we wouldn't need loras on people's face anymore with this technique
>>
File: ComfyUI_00547_.png (1.28 MB, 1024x1024)
1.28 MB
1.28 MB PNG
>>
File: applications.jpg (1.59 MB, 2477x1701)
1.59 MB
1.59 MB JPG
>>101743598
For me I hope there will be an instantID for flux, this shit is amazing I had a blast having my face on wacky pictures kek
https://instantid.github.io/
>>
>>101743424
Raw gen, for some reason Flux likes to give me such blurry images
>>
>>101743664
>for some reason Flux likes to give me such blurry images
that happens when you have a CFG > 1, dunno if you increased CFG or not anon
>>
File: ComfyUI_00550_.png (1.44 MB, 1024x1024)
1.44 MB
1.44 MB PNG
>>
File: R7OXJYf.png (63 KB, 757x546)
63 KB
63 KB PNG
https://www.reddit.com/r/StableDiffusion/comments/1ekgiw6/comment/lgpgyhf/?utm_source=share&utm_medium=web2x&context=3
>I played with a quite a few ClipTextEncodeFlux values and could not get it to adhere to the negative prompt for realistic images.
>I'll be perfectly transparent. I'm not 100% sure I fully understand difference between CFG and FluxGuidance yet, let alone ClipTextEncodeFlux.
>But as far as I know Flux uses a single guidance value learned during training, while CFG recalculates results on the fly for more flexibility and control. Why layering the FluxGuidance on top of the ClipTextEncodeFlux improves performance for me I don't have a clue
But I know was getting better results after adding in the FluxGuidance, could be a fluke but time will tell. I'll experiment more and report back if I learn anything interesting. Currently trying to integrate the negative prompting into Img2Img with Flux.
wtf? this shit is getting really complicated
>>
>>101742894
prompt?

>>101743184
>>101743250
>>101743305
>>101743378
move the monitor somewhere else (either 2nd GPU or iGPU) and it fits
--> >>101677660
>>
File: ComfyUI_00555_.png (1.3 MB, 1024x1024)
1.3 MB
1.3 MB PNG
>>
>>101743305
The 5090 will have 48 gigs bro
Trust the plan
>>
>>101743755
pre-ordering
>>
>>101743755
>Nvidia is going to cut into its enterprise GPU scam so johnny coomer can generate faster porn.

Nope, not gonna happen until China or Intel can produce a better product.
>>
>>101743755
yeah 48 gigs of RAM as a free bonus along with your 12 gb gpu
>>
File: oiu.png (830 KB, 928x1232)
830 KB
830 KB PNG
>>101743716
i made it on midjourney sorry (this too)
>>
I have a 4090 and 32GB ram. Can I run Flux at full speed?
>>
>>101743674
No CFG always 1, guidance I change depending on prompt
>>
File: ComfyUI_00564_.png (1.73 MB, 1024x1024)
1.73 MB
1.73 MB PNG
>>
>>101743663
>>101743598
It will be interesting if we didn't need to do face loras at all, jut drop in a reference pic. Looks like this should be entirely possible, as it's an img2img step
>>
>>101743755
not gonna happen, they already have a 48gb vram card and they're selling it 10 times the prices of a 3090, there's no way they're gonna let us have so much vram
>>
>>101743838
That's not really an acceptable compromise for me. I want LoRAs to do the work for me.
>>
>>101742757
what proomt?
>>
>>101743838
>It will be interesting if we didn't need to do face loras at all, jut drop in a reference pic.
it's already possible with InstantID
https://youtu.be/_8iElCxHg-c?t=495
>>
>>101743477
it wasnt easy to train, it was just the only larger model at the time that wasnt a complete meme
>>
>>101743852
>That's not really an acceptable compromise for me. I want LoRAs to do the work for me.
I don't, having to searsh for a lora for every single celebrity of anime character is exhausting, I wish we could finetune flux the shit out of it so that it knows absolutely everything
>>
>>101743852
Why? You need to drop in a Lora anyway, why not drop in a pic instead? You wouldn't rely on someone training that subject either, so it can be literally anyone, including the girl you masturbate to from class.
>>101743858
With Flux though?
>>
I'm new to Comfy and I want to upscale my Flux output image.
Will Ultimate SD Upscaler work with Flux?
What should I use?
>>
>>101743872
>With Flux though?
nope :[ but it's probably possible too, they trained a model to work on SD1.5 and SDXL, the same should be done for flux yeah
>>
>>101743872
>>101743871

When I do Loras of people or characters, I want the whole aesthetic that comes with it. That's only really possible with a LoRA or controlnet. I'm not interested in basically slapping people faces on other people.
>>
>>101743871
You will download seven thousand loras and you WILL like it.
>>
File: ComfyUI_00602_.png (931 KB, 1344x768)
931 KB
931 KB PNG
Oh god turn on CNN
>>
>>101743891
With all due intended offense you're a fucking retard
>>
>>101743891
>When I do Loras of people or characters, I want the whole aesthetic that comes with it.
but you can add any aesthetic you want into a character that is learned by the model, look at trump for example kek >>101743899
>>
File: FLUX_00038_.png (985 KB, 896x1152)
985 KB
985 KB PNG
hmm...
>>
File: [flux-dev]_00336_.png (1.17 MB, 1024x1024)
1.17 MB
1.17 MB PNG
>>101743809
>midjourney
>on Local diffusion general
heartbroken
(Thanks for responding in any case, still cool gens)
>>
>>101743899
kek'ed hard
>>
>>
File: media_GUOKt1KWMAA0V7B.jpg (120 KB, 896x1152)
120 KB
120 KB JPG
Lora works on flux, looks like the "it's impossibruu to train distilled models" meme died quickly
https://xcancel.com/ostrisai/status/1820462674230059328
>>
File: Proof of flat Earth.png (1011 KB, 1024x1024)
1011 KB
1011 KB PNG
>>101742966
>>
>>101744000
Can it be trained on a single 24gb card, do I need regularization images, especially for text and how long would it take?
>>
>>101744006
I think he trained it on SimpleTuner
https://github.com/bghira/SimpleTuner?tab=readme-ov-file#hardware-requirements
>>
https://reddit.com/r/StableDiffusion/comments/1el0rzo/announcing_dragonflye_our_new_flagship_anime_ai/
A new base model focused only on animes will be released it seems.
>>
>>101744028
but can it do 1girl big sex?
>>
File: file.png (2.51 MB, 1024x1024)
2.51 MB
2.51 MB PNG
flux is really fun. are there any other samplers than euler i should try? or somewhere to read about what the differences are?
>>
>>101744022

Flux.1 [dev, schnell]
A100-40G (LoRA, rank-16 or lower)
A100-80G (LoRA, up to rank-256)
3x A100-80G (Full tuning, DeepSpeed ZeRO 1)
1x A100-80G (Full tuning, DeepSpeed ZeRO 3)
Flux prefers being trained with multiple GPUs.

Shits dead if I can't collect a bunch of BBWs and make a BBW LoRA
>>
>>101744036
>flux is really fun. are there any other samplers than euler i should try? or somewhere to read about what the differences are?
I'll make a XY plot, do you have a prompt in mind?
>>
>>101744028
those no vae looking outputs really inspiring confidence kek
>>
>>101744036
i've been messing with different samplers.. some of them produce nothing but trash like the dpmpp_2m_sde and such, but dpm fast and ddim work good
>>
>>101744036
Probably just my imagination, but Uni PC does text better.
>>
>>101744043
I got personal loras for 1.5 and xl but I don't see it happening for flux
I'm not buying a server rig just to produce smut, I've never paid for porn and I never will
>>
File: ComfyUI_00582_.png (1.12 MB, 1024x1024)
1.12 MB
1.12 MB PNG
>>
>>101744043
negro, A100 costs peanuts per hour
>>
>>101744043
it's only a few dollars if you train it on Runpod
>>
>>101744028
I'm not a tranime fan, but it has a very pretty style, to be honest. Is it SD based or its own thing?
>>
>>101744043
I don't get people complaining about that, on /lmg/ those fuckers have to deal with fucking 70+b models, they are all training on the cloud, you can't expect flux to be good and small at the same time, this is the price to pay for having dalle at home
>>
>>101744057
>I've never paid for porn
You paid for it by buying a GPU. Training it on a rented rig is way cheaper.
>>
>>101744069
Pretty sure they run GPUs in parallel to run those large models. Can you do that with SD?
>>
File: file.png (2.63 MB, 1024x1024)
2.63 MB
2.63 MB PNG
>>101744046
these prompts basically consist of just overloading it with hundreds of elements and styles until it goes what the fuck and becomes schizo
>>
>>101744071
I also use it for video games
what good is a graphics accelerator card besides ML stuff
>>
>>101744086
wait you can't train a model in parallel on SD? I remember pony saying that he has 8xA100 running in parallel on his garage or something
>>
File: ComfyUI_00949_.png (1.29 MB, 1024x1024)
1.29 MB
1.29 MB PNG
>a 20 year old Donald Trump with dreadlocks is showing his hands
there's some areas where flux is completely clueless on what to do with the prompt
>>
>>101744028
That's not a base model, that's a SDXL finetune.
>And yes you will be able to use this with fooocus as the base model is SDXL.
Their samples look washed out (wrong VAE?) and way too 3D.
>>
>>101744115
>train
Ah, I thought you were talking about just running them. I didn't think /lmg/ actually trained anything themselves.
>>
File: FD_00153_.png (1.21 MB, 1024x1024)
1.21 MB
1.21 MB PNG
>>101744000
The irony is the impossible to train meme was started by the same guy who built the functionality into the tool used to train this LoRA.
>>
File: ComfyUI_00590_.png (1.01 MB, 1024x1024)
1.01 MB
1.01 MB PNG
>>
File: Flux_00060_.png (1.42 MB, 1024x1024)
1.42 MB
1.42 MB PNG
>>101744136
looks like Donald Trump to me
>>
>>101744141
>Ah, I thought you were talking about just running them.
you can actually put the VAE and the CLIP on another gpu with flux though, you do that the OverrideVAEDevice node, the script can be found a few threads before that one
>>
File: ComfyUI_00597_.png (1.22 MB, 1024x1024)
1.22 MB
1.22 MB PNG
>>
>>101744148
at least he reddemed himself good by showing it how possible it is
>>
>>101744156
interesting.
I'm eyeing an 8gb pascal ex-mining card for very cheap. I was gonna get it for nemo but if it can hold the T5 while my main gpu holds the model, that'll work better?
>>
File: FD_00157_.png (1.45 MB, 1024x1024)
1.45 MB
1.45 MB PNG
>>101744152
>>101744136
Sorry I just don't see race.
>>
File: [flux-dev]_00187_.png (1.29 MB, 1024x1024)
1.29 MB
1.29 MB PNG
>>101744046
nta but I thought that it was kind of you to offer
>>
>>101744149
cool prompt
>>
>>101744170
there you go I found it >>101688574
>>
>>101744136
Looks exactly like what you asked for though.
>>
>>101744178
oh sheesh, anyways I'm still waiting for a prompt, something that would showcase the difference between all the samplers for good
>>
File: ComfyUI_00602_.png (1.16 MB, 1024x1024)
1.16 MB
1.16 MB PNG
>>
File: FLUX_00043_.png (1.04 MB, 896x1152)
1.04 MB
1.04 MB PNG
>>
>>101744200
>a teenage boy instead of a 20yo man
>black skin even though I just asked dreadlocks for trump (who is a white man in case you don't remember)
>color skin aside, Trump doesn't have this nose/eyes/mouth shape at all
Yeah right... exactly what I've asked for...
>>
File: FD_00158_.png (1.72 MB, 1024x1024)
1.72 MB
1.72 MB PNG
>>101744097
>apple, banana, chair, dragon, eagle, frog, guitar, horse, island, jungle, kite, lemon, mountain, notebook, orange, piano, queen, robot, snake, tiger, umbrella, violin, whale, xylophone, yogurt, zebra, anchor, balloon, camera, dolphin, envelope, fish, grape, hat, igloo, jelly, kangaroo, lamp, monkey, nest, owl, pumpkin, quill, rainbow, ship, turtle, unicorn, vase, water, x-ray, yo-yo, acorn, blueberry, cactus, dice, elephant, feather, garlic, hammer, ink, jigsaw, key, ladder, mango, needle, octopus, pillow, quiver, ruler, star, tent, unicorn, vase, windmill, xerox, yam, zipper, ant, brick, candle, dog, ear, fan, glove, heart, ice, jug, kite, ladder, moon, nut, oar, pen, queen, ring, soap, tap, urn, vase, wave, xylophone, yam, zipper, apple, bat, car, duck, egg, fork, glass, hen, iguana, jar, kite, leaf, mouse, nail, olive, pen, quilt, rope, spoon, table, urn, vase, wheel, xylophone, yam, zebra, ant, boat, cat, drum, engine, fan, goat, hat, ink, jug, kite, lemon, map, nest, owl, paint, quilt, rose, sun, tree, umbrella, van, wind, x-ray, yam, zebra, airplane, bike, clock, dice, eagle, fan, gift, hat, ice, jam, kite, lamp, mirror, needle, orange, panda, queen, radio, star, train, umbrella, violin, worm, xylophone, yellow, zebra, apple, bike, cactus, drum, elephant, fan, goat, hat, ice, jug, kite, lemon, map, nest, owl, paint, quilt, rose, sun, tree, umbrella, van, wind, x-ray, yam, zebra, airplane, bike, clock, dice, eagle, fan, gift, hat, ice, jam, kite, lamp, mirror, needle, orange, panda, queen, radio, star, train, umbrella, violin, worm, xylophone, yellow, zebra
>>
>>101744203
you can send the prompt in the original image by uploading it to catbox right? i'll upload one of the schizo ones there in a minute.
don't laugh at the mass of schizobabble in the prompt, it's just a result of adding shit in everywhere after each generation and seeing if it makes things better or worse, endlessly.
>>
>>101744222
Checked. I think there are some baked-in negative prompts.

I absolutely can't get amputees out of Flux, thus far.
>>
File: FD_00159_.png (1.73 MB, 1024x1024)
1.73 MB
1.73 MB PNG
>>101744227
>apple, mountain, chair, tiger, lamp, balloon, orange, drum, island, kite, guitar, ocean, dragon, candle, notebook, cloud, snake, bench, eagle, pyramid, lemon, octopus, whale, ruler, jungle, clock, star, anchor, piano, bridge, violin, robot, mirror, frog, paintbrush, castle, feather, tent, pumpkin, envelope, zebra, vase, rainbow, jelly, squirrel, windmill, honey, xylophone, dragonfly, truck, quill, igloo, lamp, dolphin, pyramid, toast, kite, mountain, horse, tree, crown, hammer, book, octopus, island, needle, ship, iguana, violin, ladder, marshmallow, robot, anchor, tent, globe, apple, butterfly, fish, helmet, treasure, compass, umbrella, spaceship, rope, crystal, hat, plane, unicorn, jellybean, carrot, bat, honey, key, water, zeppelin, fish, piano, unicorn, clock, ship, jacket, cloud, mushroom, grape, whistle, glasses, elephant, peanut, scarf, tiger, swing, star, lantern, box, pillow, dragon, wolf, blanket, drum, scarf, ink, flower, cup, watermelon, fan, lighthouse, notebook, garlic, soap, hat, helmet, feather, fish, radio, cat, pumpkin, yo-yo, ring, jug, kite, airplane, river, canoe, xylophone, mountain, lizard, vase, jar, sandcastle, cloud, sailboat, mitten, galaxy, bottle, vase, elephant, spoon, worm, cactus, paint, jeep, seashell, sun, giraffe, acorn, wave, dog, olive, egg, mitten, net, seahorse, yarn, peanut, apple, marbles, basket, camel, leaf, yo-yo, fan, shell, tower, wolf, book, shoe, castle, snake, moon, cup, quilt, wind, jar, sun, volcano, bread, quiver, ear, train, ladder, camera, ink, kite, dragon, pillow, rose, hammer, quill, fish, truck, lemon, bench, pumpkin, rose, monkey, bracelet, net, arrow, cloud, kite, melon, lighthouse, unicorn, carrot, bowl, anchor, tent, grape, lamp, nut, glove, wheel, vase, ring, boat, x-ray, ball, mitten, canoe, cup, apple, drum, hat, fence, panda, ring, ladder, volcano, mushroom, bottle, quiver, pie, tomato, basket, giraffe, shell, watch, scarf, rocket, globe, mitten, tiger, dragonfly, whale, box, jelly, car,
>>
I think Heun is the best sampler for flux.
>>
>>101744227
good luck verifying everything is in there kek
>>
>>101744240
kek exactly
>>101744235
here's this, hope it works https://files.catbox.moe/j6jwx7.png
>>
>>101744235
>you can send the prompt in the original image by uploading it to catbox right?
yep
>>
>>101744212
kek
>>
>>101744241
all I know is that Heun makes things take way longer with other models. I see what you are trying to do
>>
File: ComfyUI_00605_.png (1.09 MB, 1024x1024)
1.09 MB
1.09 MB PNG
>>
File: Flux_00064_.png (1.25 MB, 1024x1024)
1.25 MB
1.25 MB PNG
>>
File: Flux_00336_.png (836 KB, 1128x904)
836 KB
836 KB PNG
Spoopy
>>
>>101744256
i been doin heunpp2 .. it takes about 2x as long as dpm but the quality seems higher
>>
>>101744261
looks perfect for eating
>>
>>101744269
ooh this is gud
>>
>>101744256
Heun mangles the text less and looks cleaner.
>>
>>101744241
deis is a sharper euler. heun is just slow
>>
File: Capture.jpg (384 KB, 2777x1474)
384 KB
384 KB JPG
>>101744136
>>101744239
cfg 3 + DynamicThresholdingFull seems to do the job fine, this is a good tool to make flux even better at prompt understanding, I wish I could find the best parameters though so that it looks less "fried" I would say
>>
File: FD_00161_.png (2.06 MB, 1024x1024)
2.06 MB
2.06 MB PNG
>>101744248
Good prompt.
>>
File: ComfyUI_00612_.png (1.13 MB, 1024x1024)
1.13 MB
1.13 MB PNG
>>
Haven't checked the state of generative models in a while. Has anyone come up with an easy way of getting consistent characters yet?
>>
File: [flux-dev]_00243_.png (1.27 MB, 1184x896)
1.27 MB
1.27 MB PNG
>>101744240
Interesting. Same seed?
>>
File: FD_00162_.png (2.07 MB, 1024x1024)
2.07 MB
2.07 MB PNG
>>101744289
Same seed, with >>101744248 slapped into chatgpt to make it verbose, then used the output as the prompt
>>
>>101744301
yeah it's called drawing
>>
>>101744301
faceswap
>>
File: FD_00160_.png (1.49 MB, 1024x1024)
1.49 MB
1.49 MB PNG
>>101744303
No different seeds.
>>
>>101744301
InstantID only works on humans unfortunately
>>
File: Flux_00339_.png (948 KB, 1128x904)
948 KB
948 KB PNG
>>
>>101744289
yeah have fun with it, step count and guidance affect it as well as the prompt. if it starts looking too "clean" and simple you need to lower guidance a bit and if it starts looking hazy and unfocused you can crank up steps, this keeps it looking good as you make the prompt more and more complex
>>
>>101744235
You literally get the full setup from the image.
>>
>>101744329
bad ass
>>
>>101744305
i like it
>>
>>101744308
Drawing is for cucks. I guess I'll have to keep waiting until a codemonkey figures how to get consistent characters then.
>>
File: Flux_00065_.png (1.19 MB, 1024x1024)
1.19 MB
1.19 MB PNG
>>
>>101744345
daaaaaamn, catbox?
>>
File: ComfyUI_00623_.png (1.09 MB, 1024x1024)
1.09 MB
1.09 MB PNG
>>
>>101744271
>>101744280
man if this is true... this + negative prompts would triple my gen times twice and become 27 mins
>>
>>101744357
wait how do you do negative prompts?
>>
File: FD_00164_.png (2.01 MB, 1024x1024)
2.01 MB
2.01 MB PNG
>>101744331
Happy with it at 3.5g and 50 steps. Gonna use that as a seet to get chatgpt to generate me a bunch of random bullshit
>>
Hit it...
>>101744342
>>101744342
>>101744342
>>
>>101744345
Can you make her an amputee?
>>
>>101744365
see >>101742498
>>
>>101744370
very nice!
>>
>>101744350
here bro
https://files.catbox.moe/dff0a2.png

>>101744380
lmao
>>
>>101744271
I got a near identical result as euler + beta and it took 3x as long
>>
Any medieval prompt? Like real medieval, not renaissance style.
>>
>>101744377
TY baker



[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.