[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 (1005 KB, 3264x3264)
1005 KB
1005 KB JPG
Discussion of free and open source text-to-image models

Previous /ldg/ bread : >>101685374

>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

>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

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

>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
>>
Flux will soon get removed, so please get it until it's too late.
>>101688274
>>101688297
>>
File: file.png (1.76 MB, 1024x1024)
1.76 MB
1.76 MB PNG
>>
blessed bread that scares SAI shills like this one >>101689069
>>
>>101689069
Sending this to every internet/AI-related news media outlet
>>
>>101689069
there is absolutely no way they're surviving located in Germany. could they turn their office into a mosque to dodge the charges?
>>
https://old.reddit.com/r/StableDiffusion/comments/1ehtpng/you_can_run_flux_slowly_on_8gb_vram/
>>
File: 1720365087125302.png (874 KB, 832x1216)
874 KB
874 KB PNG
finally a model i can use to generate decent avatars for my vidya characters
>>
>>101689112
cheeky windmill symbol of friendship
>>
Janny can you please delete this thread? We still have 170 images left on the previous one. Thanks.
>>
late blessings upon thee, young anon
>>
>>101689112
Und welche model ist das?
>>
>>101689069
Don't seethe like that Emad, you lost
>>
>>101689183
flux dev, SD could never get the armband right
>>
>>101689183
Und welches Modell ist das/es/jenes?

At least get it right.
>>
>>101689050
Give me your vrams you fucking whore
>>
Flux is getting shut down. The police are on their way.
>>
File: Lets-goo.jpg (272 KB, 2964x1398)
272 KB
272 KB JPG
I'll post this message these for those who didn't see the trick, if you have multiple GPUs, you can put CLIP on one gpu and the image model on the other gpu like this:

1) You download this ComfyBootlegOffload.py script here: https://gist.github.com/city96/

2) You put it in ComfyUI\custom_nodes then restart comfy.

I've included a workflow for those who have multiple gpu and want to to that, if cuda:1 doesn't work for you then go for cuda:0
https://files.catbox.moe/jxgi23.png
>>
>>101689112
People worry about le babies and gore when a German based company lets you make this
If they are held responsible for the gens, the swastikas will be the canary in the mine folks
>>
>Released half-assed SD3 model
>Everyone hated it
>"Lol who cares, we're the only company doing this we make the rules fuck you"
>Flux comes out
>Ohshit.png
>The SAI team is in full meltdown mode now
The germans killed SAI.
>>
File: FLUX_01524_.png (831 KB, 768x1024)
831 KB
831 KB PNG
>>
File: 1713135369155706.png (1.54 MB, 1024x1024)
1.54 MB
1.54 MB PNG
>>101689252
I genned this yesterday
>>
>>101689267
the ironic part of this story is that flux was being made by researshers who left SAI, exactly the same story as the creation of ClaudeAI (made by former OpenAI's employees)
>>
>>101689296
>ClaudeAI
sorry to nitpick but the company is called Anthropic, the model is Claude
>>
>>101689303
oh yeah my b
>>
>>101689283
Anzeige ist raus
>>
File: FLUX_01534_.png (741 KB, 768x1024)
741 KB
741 KB PNG
>>
>>101689278
>>101689311
kek'ed
>>
It's either BFL or Replicate shitting themselves, but pro is completely borked, gens hang for tens of minutes without finishing
>>
File: ComfyUI_Flux_0821.jpg (109 KB, 1024x1024)
109 KB
109 KB JPG
>>
>>101689241
Don't do this, it creates mustard gas.
>>
File: 1709337045262839.png (1.15 MB, 1024x1024)
1.15 MB
1.15 MB PNG
oy vey
>>
official pixart bigma, lumina 2 and hunyuan finetune waiting room, now with [REDACTED]
>>
>>101689338
just admit your defeat Emad, SD is history, it's flux era now
>>
File: file.png (135 KB, 256x256)
135 KB
135 KB PNG
>>101689348
now I worry about having 4archive in my dataset
you people will ruin everything
>>
>>101689323
I think there's just too many people running these now, not surprising this model is so good and the API isn't that censored compared to the competitors (dalle3, MJ)
>>
File: 1705313840604527.png (1.46 MB, 1024x1024)
1.46 MB
1.46 MB PNG
>>
File: FLUX_01545_.png (858 KB, 768x1024)
858 KB
858 KB PNG
>>101689338
it's too late for me
>>
File: test.jpg (379 KB, 1231x1642)
379 KB
379 KB JPG
>>101689278
>>101689311
lulz
>>
>>101689360
The API is completely uncensored, at least for schnell/dev since you can disable the safety checker for API requests, I can't test for pro BECAUSE ITS HANGING FOR TENS OF MINUTES AAAAAAAAAAAAAAAAAAAAAAAAAAAA.
>>
File: ComfyUI_00939_.png (776 KB, 1024x1024)
776 KB
776 KB PNG
>>101689359
CTRL+F "baby" and delete

Also HYYYYYPE
>>
>>101689380
"child" too
>>
Thread challenge: make me laugh with a funny gen. for each failed attempt i will emailing >>101688274 and >>101688297 to a random journalist
>>
File: 1722294042383680.png (298 KB, 1024x1024)
298 KB
298 KB PNG
>>101689394
>Very blurry photo of a UFO in the sky, poor quality, low quality, dark, 1980s photo, found footage
>>
>>101689380
I was like "what's the harm in having the top 20 horror movies in my dataset, the boys deserve some fun".
>>
>>101689394
what the actual FUCK is this image
>>
>>101689425
someone had a bet about whether AI could have international regulation by EOY 2024
>>
File: FLUX_01568_.png (727 KB, 768x1024)
727 KB
727 KB PNG
>>101689394
>>
Please, someone, think of the children!
>>
black tree friends will be turned into stability by the end of the year. enjoy the new safety team!
>>
File: file.png (1.06 MB, 1024x1024)
1.06 MB
1.06 MB PNG
>>101689283
>>101689112
>>
I used flux all day yesterday but I don't like it at all.
- It's slow
- Depth of field blur in every image, unable to be removed
- No negative prompting
- It's very shit for artstyles
- Girl refuses to lick banana
- It's terrible at high resolution so it really is just another 1024x1024 model
- Non-commercial license for finetuners, so it has no future
>>
File: 1720695619283278.png (688 KB, 1024x1024)
688 KB
688 KB PNG
Can someone help? I'm doing
>A 4chan inspired four-leafed green clover logo. On every leaf a Star of David with the nazi swastika inside of it.
or
>A 4chan inspired four-leafed green clover logo. On every leaf is a together merge of a Nazi Swastika inside of the Star of David.

But I can't make it generate a swastika inside of the star of david on *every* leaf.
At best it looks like picrel, in tons of cases all leafs only have star of david.
>>
>>101689485
>- It's slow
vramlet issue
>Depth of field blur in every image, unable to be removed
skill issue
>No negative prompting
cfg issue
>It's very shit for artstyles
not separately prompting clip issue
>It's terrible at high resolution so it really is just another 1024x1024 model
skill issue
>Non-commercial license for finetuners, so it has no future
reading comprehension issue
>>
>>101689487
Inpaint?
>>
>>101689512
I'm an APIbaby :((
>>
File: file.png (1.66 MB, 1024x1024)
1.66 MB
1.66 MB PNG
>>
What kind of gen times are you guys seeing with flux dev on a 4090? I'm seeing about 3 minutes. It just barely goes over 24gb.
>>
>>101689569
1024x1024 20 step fp8 4090: 15 seconds. trying to go fp16 can be 90 seconds plus thanks to the unloading
>>
File: ComfyUI_00945_.png (1.37 MB, 1024x1024)
1.37 MB
1.37 MB PNG
>>101689521
picrel

>>101689569
15 seconds on default workflow w/ fp8 all the things
>>
>>101689606
>picrel
Well it's the same model as on local, and I'm sharing prompts, it's okay right?
>>
Can Flux do Azula in a bikini?
>>
Hello. I have been out of the loop for a while (few months) and am interested in doing some image generations. Up until now I've used the Automatic1111 webui and had minimal problems with it.
When I load up SD now I get a bunch of errors in the cmd on startup. A few are just saying 'this style is deprecated' which I think are intended more for devs of extensions but I also get stuff like the following:
Error loading script: api.py
Traceback (most recent call last):
File "C:\AI ART\stable-diffusion-webui\modules\scripts.py", line 515, in load_scripts
script_module = script_loading.load_module(scriptfile.path)
File "C:\AI ART\stable-diffusion-webui\modules\script_loading.py", line 13, in load_module
module_spec.loader.exec_module(module)
File "<frozen importlib._bootstrap_external>", line 883, in exec_module
File "<frozen importlib._bootstrap>", line 241, in _call_with_frames_removed
File "C:\AI ART\stable-diffusion-webui\extensions\sd-webui-controlnet\scripts\api.py", line 90, in <module>
ControlNetTxt2ImgRequest = create_controlnet_request_model(StableDiffusionTxt2ImgProcessingAPI)
File "C:\AI ART\stable-diffusion-webui\extensions\sd-webui-controlnet\scripts\api.py", line 81, in create_controlnet_request_model
'controlnet_units': (List[ControlNetUnitRequest], Field(default=[], docs_default=[ControlNetUnitRequest()], description="ControlNet Processing Units")),
NameError: name 'List' is not defined

The webui still starts up and runs fine but I'm wondering if I am losing out on functionality here. I'm a retard who relies on retard-friendly-guides and would appreciate (1) knowing what if anything these errors do/mean and (2) how to fix them.

Also, a few months ago I messed around briefly with ComfyUI and SDXL but found that I was happier with the models and Lora that I already had. Is there a significant reason to switch? is this Flux model I hear about compatible with SD1.5?
>>
>>101689599
ok I need to figure out what I'm doing wrong
>>
File: ComfyUI_00947_.png (1.2 MB, 1024x1024)
1.2 MB
1.2 MB PNG
>>101689612
>>
>>101689653
Do you have anything else running in the background chewing up vram?
>>
File: _KvF3OZBFXah9QtII0Sju.png (798 KB, 768x1024)
798 KB
798 KB PNG
Flux did a good job.
Gen is just base model, not bad.
Will be downloading before it's deleted.
>>
>>101689679
Don't think so, using the fp8 dev model, also tried the smaller clip model.
>>
>Flux can't do minions getting high on fart gas
it's over...
>>
>>101689754
how far we have fallen
>>
>flux is gonna get baleeted!!1!1!1
this is just a meme....... right?
>>
File: file.png (1.51 MB, 1024x1024)
1.51 MB
1.51 MB PNG
>Homelander from the Boys beating a african american to death. A woman masturbates in the background
NOT WHAT I ASKED FOR FLUX
>>
now that the dust has settled, what's the official /ldg/ verdict on flux?
>>
>>101689786
Dalle at home but finetuning ability is up for debate. Hopefully its possible / they are fair with their commercial licenses for finetuners.
>>
File: ComfyUI_00951_.png (1.31 MB, 1024x1024)
1.31 MB
1.31 MB PNG
>>101689786
>>
File: ComfyUI_Flux_0859.jpg (99 KB, 1024x1024)
99 KB
99 KB JPG
>>
>>101689786
No Laura Kinney boobies
>>
>>101689786
The model has very nice comprehension, exceeding Dall-e in some areas and lagging in a few others. The model's stylistic rendering is quite rigid and bland, lacking the ability to do gritty art like Midjourney. The pop-culture knowledge is also a bit basic. Overall it's far better than the previous local alternatives and about where SD3 should've been if Emad wasn't a liar. I don't foresee this model getting a mass adoption from finetunes but rather marking the turning point for local models where they finally start taking things seriously. If hardware wasn't an issue the finetunes would be insane, but I remain doubtful and we've had plenty of models float by without anything getting developed for them.
>>
>>101689786
let me run it sub 8g vram and only then will it be good
>links to fp8
yeah, ill try it later
>>
File: miku booba.png (1.03 MB, 768x1216)
1.03 MB
1.03 MB PNG
booba
>>
File: ComfyUI_01559_.png (1.24 MB, 1152x896)
1.24 MB
1.24 MB PNG
>>101689786
>>
>>101689786
>now that the dust has settled
It has? Model came out yesterday kek
>>
>>101689877
rips top off and starts sucking on tits
>>
>>101689888
stop, it's too hot out to get horny rn
>>
>>101689786
current SOTA model for avatarfagging
>>
the enshittification of ldg has begun
>>
File: ComfyUI_00953_.png (1.01 MB, 1024x1024)
1.01 MB
1.01 MB PNG
>>
>>101689927
we should have predicted this the moment they started posting mikus
>>
File: file.png (2.67 MB, 1024x1024)
2.67 MB
2.67 MB PNG
>>101689839
i used this post as prompt
pic related is the result
>>
File: ComfyUI_00954_.png (691 KB, 1024x1024)
691 KB
691 KB PNG
>>
>>101689957
its like someone took 1.5 revanimate cardos slop and made a shrunken head out of it
>>
File: file.png (1.5 MB, 1024x1024)
1.5 MB
1.5 MB PNG
>>101689957
and this is what base pixart sigma gave. at the end the pixartsexuals still run /ldg/
>>
>>101689957
>>101689994
pixart mogs
>>
Flux fatties.. The pixshartists are making fun of us with their micro models again.. Our response?
>>
>>101689930
Hmm, could use Flux to gen comic cover with proper text and sfw waifu, then take image into sd1.5 or pony inpaint and slut up the waifu
>>
is there a blingee gif model, ideally just load an image dial in level of bling on a slider, and the model decides which sub-blings to use and where to put them
>>
im telling you, flux just aint all that. bigma can do what flux does with half the params.
>>
>>101689786
2MW until shit makes it usable, just like what happened with SD
>>
File: 241124234457568.png (646 KB, 1473x729)
646 KB
646 KB PNG
works on cpu at "usable" speeds
>>
>>101690109
>pixart text
>>
>>101690123
acting like the average imagefag has the mental capacity to read
>>
>>101690121
just use the api doebeit? it's 0.3 cents per image and takes 1 second
>>
File: file.png (1.06 MB, 1024x1024)
1.06 MB
1.06 MB PNG
Sissies... it's over... it doesn't work...
>>
>>101690198
Prompt?
>>
File: a2_b5_cfg1.4.png (1.51 MB, 768x1344)
1.51 MB
1.51 MB PNG
Am I missing something with CLIPTextEncodeFlux? I get that it needs cfg greater than 1 in the sampler, but that just wrecks gen time. Negative prompting is nice though.
>>
File: ComfyUI_temp_ppftb_00200_.png (764 KB, 1024x1024)
764 KB
764 KB PNG
>>101690109
>>
File: ComfyUI_01564_.png (1.46 MB, 1152x896)
1.46 MB
1.46 MB PNG
>>101689856
>>
File: file.png (1.3 MB, 1024x1024)
1.3 MB
1.3 MB PNG
>>101690209
>A political comic in three panels: Panel one on the top left shows a man and a woman. The woman is saying "Men only want one thing and it's disgusting". Panel two on the top right is just the man, saying "Actually I'm gay." Panel three on the bottom is just the woman, with a look of anger.
I just want to test if I can re-make shitty StoneToss comics with it but it's struggling.
Maybe fp8 is too much of a detriment.
>>
>>101690318
still kekd at it
>>
File: ComfyUI_Flux_0919.jpg (104 KB, 1024x1024)
104 KB
104 KB JPG
>>
File: 1722531043184238.png (1.05 MB, 1024x1024)
1.05 MB
1.05 MB PNG
Does FLUX run on Automatic11111 yet? Too much of a brainlet to use ComfyUI
>>
>>101690318
flux dev on replicate 6 gens with the prompt, no dice...
https://files.catbox.moe/7iussd.jpg
https://files.catbox.moe/9srah1.jpg
https://files.catbox.moe/g2z2sl.jpg
https://files.catbox.moe/5qto2z.jpg
https://files.catbox.moe/1yzl9g.jpg
https://files.catbox.moe/m3boy2.jpg
>>
File: ComfyUI_Flux_0187.jpg (92 KB, 1216x832)
92 KB
92 KB JPG
>>101690361
>>
Pixart can't into memes it's so fucking over
>>
>>101690361
Have to wait few days until hetero bros can use it
>>
>>101690361
>brainlet
you just download the three models, put them in the folder, and then steal someone elses catbox to copy their setup.
>>
>>101690386
reminds we of what i said about sd last october when dall-e dropped, and all the sdkeks had an absolute shitfit ahahaha
>>
remember when sigma was "too hard" to install? same process for flux kek
>>
File: 1695459007814995.png (1.21 MB, 1024x1024)
1.21 MB
1.21 MB PNG
NOOOOOOOOOOOOOOOOOO IT COULD'VE WORKED NOOOOOOOOOOOOOOOOOOOOO
>>
File: 1701443364891499.png (1.13 MB, 1024x1024)
1.13 MB
1.13 MB PNG
kind of?
>>
File: 1710952024326741.png (683 KB, 1024x1024)
683 KB
683 KB PNG
got it, wrong layout but it worked! sadly small mistake in text - no "one"
>>
File: 1711100680043286.png (870 KB, 1024x1024)
870 KB
870 KB PNG
I'm just gachaing the system by doing tens of requests to replicate dev, who needs inpaint when you can gacha
>>
File: 1721920367104763.png (1.06 MB, 1024x1024)
1.06 MB
1.06 MB PNG
>>
>>101690318
>>101690555
>>
>>101690259
kek
>>
File: file.png (1.3 MB, 1024x1024)
1.3 MB
1.3 MB PNG
>>
File: 00137-321881170.jpg (1.27 MB, 1568x2016)
1.27 MB
1.27 MB JPG
flux for coom when?
>>
>>101690630
never, sorry, go back to sleep
>>
Does anyone know how to get Matrix style green 1s and 0s raining down in ghe background?
>>
File: ComfyUI_04196_.png (1.16 MB, 1024x1024)
1.16 MB
1.16 MB PNG
>A group of mischievous Minions with oversized, comically large feet. They are in their usual yellow attire with blue overalls, but their feet are exaggeratedly big compared to the rest of their bodies. They should have a playful and silly expression, and the scene should be vibrant and animated, capturing their quirky nature
erm..... minionbros..?
>>
File: ComfyUI_Flux_0953.jpg (157 KB, 1024x1024)
157 KB
157 KB JPG
>>
File: dalle minions.jfif.jpg (119 KB, 1024x1024)
119 KB
119 KB JPG
>>101690751
dalle mogs here....
>>
File: bgFcWLrYENK77_6bfTo53.png (997 KB, 768x1024)
997 KB
997 KB PNG
>>101689786
Not bad, feels closer to what Dall-e can achieve with how it interprets the prompts but still very much SD.
>>
File: file.png (1.65 MB, 1024x1024)
1.65 MB
1.65 MB PNG
>specify crude and poorly drawn in shitty pencil
>it's immaculate
It's trash.
>>
File: file.png (2.24 MB, 1024x1024)
2.24 MB
2.24 MB PNG
>>101690751
Skill issue?
>>
>>101690799
You don't understand how T5 works. Its very literal, like talking to a dictionary. Tell it that the drawing is messy / the lines are messy
>>
File: ComfyUI_Flux_0957.jpg (158 KB, 1344x768)
158 KB
158 KB JPG
>>
>>101690783
it can do overlapping rotated Z?
>>
>>101690828
who made it like this? for what purpose?
>>
File: file.png (1.9 MB, 1024x1024)
1.9 MB
1.9 MB PNG
>>
>>101690845
Are you retarded?
>>
>>101690845
blind person with big brain vs autistic child but can point and say "that's booba!"
>>
>>101690839
Fantastic, will Flux do stuff like murders, horror genre? I have ideo because it's less bitchy (still anti-adult-nudity, which is retarded)
>>
>>101690799
try "a child's simple crayon drawing of"
>a child's simple crayon drawing of a house
>>
>>101690891
yeah >>101688274 >>101688297
>>
if it can't extrapolate messy lines from 'poorly drawn' then it's a complete fucking failure. it isn't "a dictionary", it's retarded.
>>
>>101690891
*ideogram

it sucks at hands still.
>>
>>101690904
lol
>>
File: file.png (1.48 MB, 1024x1024)
1.48 MB
1.48 MB PNG
>>101690828
I did and it just made Sonic himself crude but still not sloppily drawn.
>>
>>101690397
and enjoy your virus
>>
File: ComfyUI_00971_.png (1.59 MB, 1024x1024)
1.59 MB
1.59 MB PNG
>>
File: file.png (1.57 MB, 2182x972)
1.57 MB
1.57 MB PNG
>>101690768
llava does not recognize it properly
>>
>>101689856
bush?
>>
File: 00154-2673513437.jpg (1.23 MB, 1568x2016)
1.23 MB
1.23 MB JPG
>>101690688
am disappoint
>>
>>101690975
t: hatsude gigu
>>
reminder to turn guidance (not cfg) down to 2.5, especially if you're generating art
greatly reduces the "slopped" look compared to 3.5 and gives you better hand drawn/painted textures

it's better for photos too though imo, just greatly reduces the slop/overcooked look in general. the default 3.5 is like having the cfg cranked up to 9 on stable diffusion
>>
File: ComfyUI_00094_.png (1.4 MB, 1024x1024)
1.4 MB
1.4 MB PNG
>>101690630
>>101690688
2nd ever flux gen
>>
flux has deformed feet.

Pathetic.
>>
File: ComfyUI_Flux_0965.jpg (125 KB, 1152x864)
125 KB
125 KB JPG
>>101691048
turned it down to 2. also using PAG but i'm not sure if it does anything for flux and too lazy to test it lel
>>
File: file.jpg (108 KB, 1024x1024)
108 KB
108 KB JPG
>>101690897
No dice. He's still too clean.
>>
>>101691048
the nodes dont seem to have guidance? at least the default comfy workflow doesnt. is there a better one available?
>>
>>101691085
try "drawled by a retarded"
>>
File: dalle sonic fail.png (1.17 MB, 1309x828)
1.17 MB
1.17 MB PNG
>>101691085
i tried this on dall-e and midjourney and it gets similar results on both. are you able to get it out of any other model?
>>
File: ComfyUI_Flux_0967.jpg (121 KB, 1152x864)
121 KB
121 KB JPG
>>
>>101691089
comfy only patched the guidance node in a few hours ago, update your UI and then add the new node FluxGuidance node and connect it between the prompt box and the sampler
>>
File: Sonic's House.png (1.17 MB, 768x1216)
1.17 MB
1.17 MB PNG
>>101690923
>>101691085
It just flat out refuses to draw Sonic shittily.
>>
>>101691121
Help a promptlet out
>>
File: ComfyUI_temp_ppftb_00270_.png (2.14 MB, 1024x1024)
2.14 MB
2.14 MB PNG
>>101690751
>minion feet
>>
File: 00000-1983771864_cleanup.jpg (3.11 MB, 2352x3024)
3.11 MB
3.11 MB JPG
>>101691130
>prompting sonic instead of sanic
>>
>>101691174
PROOOOOOOOOOMPT?
>>
Can Flux gen this https://www.youtube.com/watch?v=E0ZHXVp_wUE
>>
File: unknown (8).png (1.5 MB, 1024x1024)
1.5 MB
1.5 MB PNG
>>101691157
it works fine in dalle but flux seems to hesitate at genning monstrous feet for them
>>
File: file.png (1.21 MB, 1024x1024)
1.21 MB
1.21 MB PNG
>>101691174
Sanic gives adjacent hedgehogs.
>>
>>101691201
>spare toes

When will the maestro of the toe appear and save us all?
>>
File: ComfyUI_Flux_0963.jpg (161 KB, 1024x1024)
161 KB
161 KB JPG
>>101691152

most of it is a dark souls character description
https://darksouls.fandom.com/wiki/Gwyn,_Lord_of_Cinder#Description

https://files.catbox.moe/wliuo8.webp
>>
>>101691229
put in drawled by a retarded
>>
>>101691255
I did in that one. It has no effect.
>>
Why do all these models use T5 as the text encoder? Why not use a more modern, autoregressive LLM like llama 3 8b (or even mistral 7b if llama 3 was too recent)?
>>
>>101691271
what the hell are you talking about you god damn ape
>>
>>101691271
because someone already made t5 for them, and image gen model baking consists of bashing together components and hoping it works
>>
flux can't do retards, trisomy etc. FAIL

GIVE ME LIBERTY OR GIVE ME DEATH
>>
>>101691271
I believe it is because for image generation purpose, the language model used has to have an encoder-decoder architecture (I don't know why)

while modern language models like llama are all decoder only
>>
>>101691284
What part of my comment was unclear you retarded gorilla nigger? When you want to do LLM stuff with a small model, you reach for mistral or llama, not fucking T5. So if you only need to use the language model's embeddings as general purpose text conditioning, seems like these other models would be better.
>>
>>101691271
What does Copilot do? It may be possible to bypass the encoder phase?
>>
File: ComfyUI_Flux_0979.jpg (93 KB, 1152x864)
93 KB
93 KB JPG
>>
>>101689241
But will this work if one GPU is AMD?
>>
Guys help spiders are burrowing into my arm and stealing all my arm jam.
>>
>>101691316
thats what the last hidden state is for...
>>
>>101691340
Can you tell it to place the figure in a stone archway, and specify the style of it?
>>
>>101691324
>git pull -> coom
this is the workflow. what the hell are you talking about
>>
>hurr why no llama
maybe because that would mean having to load another 12+GB model unquantized because the land of unet-based projects fucking sucks when it comes to actual efficiency
>>
>>101691048
I am not seeing this, depending on the subject it won't use a style at all and give you a photograph. I think it's best for the time being to img2img to a SDXL based model.
>>
File: 00002-3630594880_cleanup.jpg (2.5 MB, 2352x3024)
2.5 MB
2.5 MB JPG
>>101691186
it's just autismmix
>score_9, score_8_up, score_7_up, BREAK,(realistic:1.4), 1girl, souryuu asuka langley, plugsuit, petite, skinny, flat chest, hands on hips, (view from below:1.2), tsundere, angry, looking at viewer, (ass:1.2)
>>
Why can't amd make a gpu that has a crazy amount of ram? Why is that hard to do? Like give us literally 128gb of ram on a 7600 xt, why not?
>>
>>101691064
share the sauce
>>
>>101691271
lumina
>>
>>101691229
>>101691229
stop putting "eric age 11" I have an 11 year old and he can draw competently like this
>>
>>101691399
It will drive the ram price down, not gonna happen
>>
>>101691431
he cheated, you are stupid, I can't believe you fell for Eric and his lies.
>>
File: Sonic's House.png (1.4 MB, 768x1216)
1.4 MB
1.4 MB PNG
After a dozen gens, I kind of got it to draw Sonic badly but not as much as I would like to.
>>
>>101691457
can't you just replace original noise with your own faggot drawings so it will be properly shitty
>>
>>101689050
God bless Flux, now that we got Dalle tier I'm still waiting for a competent team to drop a model on par with Udio.

Also, it can finally handle my
>Girl squatting on top of X
coomer prompts so I can get a first person view, nice!
>>
>>101691457
sovl
>>
>>101691505
What I want is a music one that does parodies.
>>
>>101691411
Ah, I see, that uses Gemma-2b. So I guess there's no technical reason why an autoregressive LLM wouldn't work.

Would be interesting to see a future model try using Mistral-Nemo. 12b parameters, apache 2, base model exists, very uncensored and broad knowledge base, and widely considered to punch well above its weight.
>>
>>101691505
That is, before the powers that be get to them. AI music is very important.
>>
>>101691457
hopefully there'll be an IPAdapter made for Flux eventually. ipadapter is perfect for communicating a desired art style like this, you'd just input an existing shitty crayon drawing of the kind you're looking for to show the model what you mean.
>>
>>101691530
>autoregressive LLM
a what?
>>
>>101691544
flux is obviously boosting the text in the background, significantly.
>>
>>101689241
you're a saint, thanks anon

also, 3090 + 3060 chads rise up. there's so many of us
>>
>>101691530
meta's chameleon
https://github.com/GAIR-NLP/anole
>>
File: 0dqdvsdrg3gd1.jpg (385 KB, 1024x1024)
385 KB
385 KB JPG
>>101691399
they dont want to make any effort, only cruise just below nvidia and undercut them by $100 while doing the bare minimum
>>
>>101691552
Predicts the next word, one at a time, rather than something more complex like encoder-decoder, masked language modeling, sequence to sequence bullshit etc. Basically all modern LLMs that generate text are decoder-only autoregressive models.
>>
>>101691408
https://files.catbox.moe/g3o9jq.png
I think it was luck
>>
>>101691642
Run in pony inpaint and just inpaint some nipples
>>
>>101691642
It can kind of do pale ghost nipples sometimes but it's not because it's necessary good at it. They become terrible the moment you try to add any color to them.
>>
>>101691617
Did you do Dusk?
>>
File: 00161-3062146756.jpg (1.23 MB, 1568x2016)
1.23 MB
1.23 MB JPG
what the hell happened to this place
>>
>>101691399
This is what makes me think Nvidia/AMD/Intel are operating as an informal cartel. Any one of them could gain market share over the others at any time by releasing a card with a large amount of slow VRAM and mediocre cores for home AI tinkering that wouldn't be very attractive for datacenter use or gaming. But none of them do. There's no explanation for that other than some kind of cartel arrangement.
>>
I mostly use SD for creating loras of real people (celebrities) and then creating erotic images of them. What models are best for photorealistic and also NSFW generations?
>>
>>101691457
img2img over sonic with SD
>>
>>101691680
>oh oops we have a recall too, so you can get your real one fixed
>>
File: file.png (935 KB, 1024x1024)
935 KB
935 KB PNG
>>101690845
It's always been this way. You do realize how image models work right? It doesn't think like a person, it associates captions with clumps of color. It does not "think". It has never "thought". SD 1.4 basically worked by typing in image alt tags, that's how you prompted it. It has little to no ability to infer context and similarities so you must understand the CAPTIONS it was trained on. That's of course assuming they even trained on children's drawings and that those drawings were correctly captioned.
>>
>>101691399
h100
>>
>>101691505
>>101691528
>>101691543
I want a model that I can finetune on my fav band discography to get new songs
>>
>>101691717
Another idea is "Nirvana, sing the news"
>>
I don't know, the whole idea of putting the style in CLIP and the rest in T5 doesn't seem to accomplish much for me. I even tried the workflow from a previous thread but it doesn't help at all. And I know CLIP knows Jean Leon Gerome's style.
>>
>>101691680
i dunno, nvidia is currentmy crushing them too much on discrete gpus, amd/intel got fucked in the arrangment
>>
>>101691705
>eyelashes
Die.
>>
>>101691680
The problem is you have customers willing to pay $$$ for GPUs and you have poorfags. If consumer cards get too good the paypigs buy the consumer cards.
>>
>>101689241
absolutely gigabased. just being able to offload the VAE to the other card solved the problem I was having with the system slowing to a crawl when running flux in fp16 on the 3090

no more fp8, all my homies hate fp8. it's crazy how much of a difference it makes to image quality, 8bit isn't lossless for image models at all like it is for LLMs
>>
File: ComfyUI_Flux_1013.jpg (185 KB, 1152x864)
185 KB
185 KB JPG
>>
>>101691789
That's why I specified that it would be slow, cheap vram and cores, not fast like datacenters want

corpos would have no interest in a card with high VRAM but the speed of a 2080, if they did they would never have moved on from their P40s
>>
>>
I managed to do the q-flip with semi-regressive U-pole snatch. It just needs to be loaded with fp16
>>
File: FD_00011_.png (1 MB, 1024x1024)
1 MB
1 MB PNG
>>101691457
>a 4 year old childs crayon drawing of a blue hedgehog man with red shoes
>>
File: ComfyUI_Flux_1009.jpg (128 KB, 1152x864)
128 KB
128 KB JPG
>>101691837
hell yea
>>
File: FD_00013_.png (1.08 MB, 1024x1024)
1.08 MB
1.08 MB PNG
>>101691457
>>101691858
>a childs crayon drawing of a blue hedgehog man with red shoes
>>
File: ComfyUI_Flux_0785.jpg (95 KB, 1024x1024)
95 KB
95 KB JPG
>>101691837
>>
File: FD_00015_.png (948 KB, 1024x1024)
948 KB
948 KB PNG
>>101691457
>>101691858
>>101691874
>a childs poorly drawn crayon drawing of a blue hedgehog man with red shoes
>>
File: ComfyUI_temp_ppftb_00324_.png (2.14 MB, 1024x1024)
2.14 MB
2.14 MB PNG
>>101691201
Yeah, Flux struggles with feet from certain angles, especially the soles, but it's often surprisingly good, rarely missing digits or nails from top down view. Also the bigger the model, the easier it typically learns, so footfags should get their finetunes soon enough.
>>
File: 1691668688231773.jpg (247 KB, 1024x1024)
247 KB
247 KB JPG
>>101691858
This is still much worse than DALL-E 3 with Natural style, sadly. See picrel and catboxes:

https://files.catbox.moe/t3ydjy.jpg
https://files.catbox.moe/vu0hms.jpg
https://files.catbox.moe/mbmewp.jpg
https://files.catbox.moe/emi45x.jpg
https://files.catbox.moe/qms3es.jpg
https://files.catbox.moe/ftob0v.jpg

Prompts were these two:
>low-quality childish drawing of sonic on paper in pencil, straight lines
>low-quality childish drawing of sanic sonic on paper in pencil, straight lines

These prompts were used with JB of course, so it didn't rewrite them.
>>
>>101691838
Based
>>
>>101691823
You're not going to buy a 2080 with 24 GB of VRAM for $800.
>>
File: 1692761065819255.png (1.52 MB, 1024x1024)
1.52 MB
1.52 MB PNG
>>101691914
>https://files.catbox.moe/qms3es.jpg
holy sovl...
>>
>>101691945
The problem solves itself anyways, 3090s are $700 used.
>>
>>101691945
I don't think you understand how cheap VRAM can get when it's not very fast
It's not really expensive at all
>>
can someone pls leak dalle3 it can't be bigger than flux surely pretty thanks
>>
>>101691972
If I put a 3090 in my extra pci slot, will it screw up my gaming with my 6950?
>>
>>101691914
>>
Fellow 24GB gpu poors, what's the meta lads? FP16 somehow? I can generate the first image in ComfyUI lowvram mode but the second one would OOM for some reason
>>
>>101692003
I don't think AMD and Nvidia play nice in the same system.
>>
>>101692009
I said Natural DALL-E 3 style, Anon. You can only use that on the API and IIRC ChatGPT Plus if you ask it explicitly. Bing creator always uses Vivid.
>>
File: ComfyUI_00054_.png (806 KB, 768x1216)
806 KB
806 KB PNG
>>
Lack of negative prompts is fucking brutal. I'm certain "photo, bokeh" would be enough to get my desired aesthetic.
>>
File: ComfyUI_00975_.png (1.1 MB, 1024x1024)
1.1 MB
1.1 MB PNG
>>
File: FD_00024_.png (1.15 MB, 1024x1024)
1.15 MB
1.15 MB PNG
>>101692017
>this local model doesn't perform as well as as this SaaS model that can only run on A100s
>>
>>101692086
we don't know about dalle3 size
>>
File: ComfyUI_00976_.png (1.13 MB, 1024x1024)
1.13 MB
1.13 MB PNG
>>
>>101692090
An educated guess is it can do batch size 4 on an 80GB card.
>>
>>101691945
no but $800 for 2080 with 64GB yes
>>
>>101692079
>shieet muh niggums
>>
File: FD_00025_.png (1.45 MB, 1024x1024)
1.45 MB
1.45 MB PNG
>>101692090
I do know that Dalle3 can't make massive knockers like this without dogging me
>>
ideogram refuses to make crude drawings.
>>
>>101691625
>>101691552
Autoregressive means that the next token is predicted based on the the past output + prompt. If t_n is the token to be estimated, then P(t_n | t_{n-1}, t_{n-2},...,t_0, prompt) is the probability distribution of your next token. It has nothing to do with the network architecture. The issue there is that autoregressive models are, by definition, unable to plan, because they only look at the past.
>>
>>101692114
Give me the prompt and I'll make you some
>>
>>101692106
hahahahaha never in a million years poor fag
go work at mcdonalds and save up some paychecks
>>
>>101692079
Did you prompt for black persons
>>
>>101692124
Do literally anything sexual or racist and post how many times you got dogged.
>>
File: ComfyUI_00977_.png (1.08 MB, 1024x1024)
1.08 MB
1.08 MB PNG
>>101692110
>>101692128
>>
>>101692166
I have azure endpoints with tweaked filters, anon, I don't use the slop retarded vivid style bing creator/designer. Those endpoints have basically all checks disabled except for some prompt filtering, specifically there's no NSFW output image check.
>>
File: 00007-1977344558_cleanup.jpg (3.1 MB, 2352x3024)
3.1 MB
3.1 MB JPG
>>
>>101692178
give your best nsfw generation of catwoman in the next 60 seconds
>>
>>101692127
that's the point that anon was making
any one of the three COULD produce such a card using cheap slow vram and gain marketshare from others
but (as you are implying) they will not, because they are operating as an informal cartel
>>
>>101692178
If as you say you exclusively use azure API for imagery, then why are you on the LOCAL diffusion general?
>>
>>101692174
shalom
>>
>>101692174
>>
>>101692191
i never genned her before, dalle3 isn't that fast anon. wait a bit
>>101692214
Where did I say that I ONLY use Azure? I just use what's best, currently I like Flux too.
>>
>>101692197
Your stupid ass card idea only appeals to prosumers and businesses you moron. And those people can afford $2000+ for a GPU. No one wants a fucking 10s/token 2080 64 GB card. It's a dumb idea that is crafted narcissistically.
>>
File: file.jpg (203 KB, 1280x1024)
203 KB
203 KB JPG
>>
>>101691790
the appeal is that an fp8 12B model is going to outperform an fp16 6B model while using about the same amount of vram, not that it's free perfomance. it just so happens that we don't have any 6B models to compare it to, and there are other factors that affect model quality.
>>
>>101692232
no, the false idea that such a card would have any appeal to business/commercial customers would already addressed upthread
>>
File: FD_00033_.png (1.44 MB, 1024x1024)
1.44 MB
1.44 MB PNG
>>101692178
>except for some prompt filtering
>>
>>101692191
dalle3 struggles with nipples as much as flux, wait a bit https://files.catbox.moe/2ufceh.jpg
>>
>>101692174
Havent seen a debo like this since I had diarrhea
>>
>>101692250
The only people who want a 64 GB VRAM card are prosumers you fucking retard. At this point you just sound like a 12 year old that asks "why don't they give ME a ferrari for $5000"
>>
>>101692232
>No one wants a fucking 10s/token 2080 64 GB card
NTA but I unironically want this. So would half of /lmg/. You don't know what you're talking about.
That would still be orders of magnitude faster than running big models with partial CPU offloading like we have to do now.
>>
>>101690383
coincidentally, it will also take that long for furries to fix flux
>>
>>101689241
Awesome, do you know how much vram it saves from the gpu doing image gen?
Or asked another way, how much vram clip uses?
>>
>>101692270
No you don't and you would save more time getting, you know, a job.
>>
File: file.jpg (110 KB, 1280x1024)
110 KB
110 KB JPG
>>
>>101692260
https://files.catbox.moe/jrblxf.jpg yeah can't be bothered
>>
>>101692268
the 'ferrari' analogy does not work at all because as I have said 3 times now, such a card would not be expensive to produce. so analogizing it to a luxury car makes no sense
you're just wildly spitting bile without paying any attention to what's being said or bothering to make sense
>>
>>101692288
I accept your admission of defeat.
>>
>>101692232
/lmg/ would want those. They're unironically buying Macs because of their gigantic hybrid RAM which is actually slower than a 3060
>>
>>101692294
That card would be expensive to produce because someone has to work on it so they can sell 1000 units to fags on 4chan. But if it's so easy, I look forward to your homebrew 2080 with 64 GB of VRAM.
>>
File: ComfyUI_00980_.png (852 KB, 1024x1024)
852 KB
852 KB PNG
>>101692264
>>
>>101692292
skibidi is the one brainrot bullshit I don't hate zoomers for
>>
>>101689241
Do you know if this can be added into comfyui ?
I don't want that to break next time I update.
>>
>>101692315
And... how much are those Macs? Are they cheaper than A6000s?
>>
>>101692316
This is very true. Remember when Linux was so niche so no hardware vendor made any drivers for it?
This is where local SD is right now. We are not even a blip on their radar
>>
>>101692329
It's not Zoomers it's gen A.
Anyway we watched some retarded fucking shit back in the day too. https://www.youtube.com/watch?v=vIvtVSzKj8c
>>
>>101692314
Missing a picture of interior
>>
>>101692342
192GB for the same price as an A6000 (48GB), lower energy usage, more portable and less headache
>>
>>101692342
What does that have to do with him absolutely obliterating your bullshit claim that people aren't interested in large amounts of RAM combined with weak/slow compute?
>>
File: ComfyUI_00981_.png (1.09 MB, 1024x1024)
1.09 MB
1.09 MB PNG
>>
>>101692380
>"nogens"
fuck off back to /sdg/, debo
>>
>>
File: file.jpg (172 KB, 1280x768)
172 KB
172 KB JPG
>>101692329
True. It REALLY activates the brainrot in my mind.
>>
>>101692374
>bro they're going to make so much money selling 1000 GPUs to faggots on 4chan
>bro they'll just make them really fucking slow so it only appeals to autistic man children that are willing to wait 10 seconds per token so they can masturbate to their AI girlfriends
>of course this target demographic is extremely poor so you need to hit a $400 price point or they'll call you a jew
I hate this website so much
>>
>>101692397
you can just tell by his absolute dogshit prompts. new cool model comes out and all he can think of is generating text with it.
>>
File: FD_00040_.png (1.43 MB, 1024x1024)
1.43 MB
1.43 MB PNG
>>
>>101692174
I didn't mean it like that, but all people where black in one pic, and all white in the other
>>
>>101692414
you're kinda bringing it on yourself by constantly shifting your arguments, contradicting yourself and making yourself look incoherent
there's a good version of your argument to be made but you're not making it because you're too lazy to follow the conversation
>>
File: ComfyUI_Flux_1061.jpg (119 KB, 1152x864)
119 KB
119 KB JPG
>>
File: file.jpg (90 KB, 1280x768)
90 KB
90 KB JPG
>prompt leaks as text
No... it's over...
>>
File: file.png (1.1 MB, 1024x1024)
1.1 MB
1.1 MB PNG
>>101692371
>$9,599.00
>>
Flux challenge: gen someone flipping off the camera. So far I haven't been able to.
>>
>>101692288
a second job wont get me a 64GB card
I already have a 3090 there's nothing else for me over 24GB
>>
>>101692470
>cp/ in prompt
anon you are going to prison
>>
>>101692452
They will not make money selling GPUs to you. So try better than arguing like a 14 year old socialist.
>>
>>101692316
>I look forward to your homebrew 2080 with 64 GB of VRAM
that's exactly what people are trying n this very thread with 3090+3060 offloading
>>
>>101692496
I heard you can use TWO 3090s. It's crazy.
>>
https://2080ti22g.com/
>>
>>
>>101689498
>not separately prompting clip issue
What is this anon?
>>
i can't get flux dev to actually output normal looking nudity. the nipples are all blobby.
>>
File: ComfyUI_Flux_01849_.png (976 KB, 1024x1024)
976 KB
976 KB PNG
The feet look a little airburshed but it can do them
>>
>>
>>101692517
why are they making 4070 / 4080 / 4090s then?
>>
>>101692533
so close... now count on your fingers how much 24+24 is compared to 64
>>
File: ComfyUI_00986_.png (873 KB, 1024x1024)
873 KB
873 KB PNG
>>101692421
Why can't I be a promptlet and not debo

>>101692435
Oh true. Interesting observation idk
>>
>>101692592
They aren't going to sell you a 64 GB card incel.
>>
>>101692491
You can customize to go for less than 5k, you don't need 8TB of storage. That's about the same price as the A6000
>>
>>101692574
Because they have target audiences and volume?
>>
File: FD_00046_.png (1.61 MB, 1024x1024)
1.61 MB
1.61 MB PNG
>>101692602
>>101692602
Don't worry Anon I can tell you aren't debo because you don't have an extremely obvious file name.
>>
File: file.png (62 KB, 522x970)
62 KB
62 KB PNG
>>101692620
stop lying
>>
File: ComfyUI_Flux__00006_.png (788 KB, 1024x1024)
788 KB
788 KB PNG
>>
>>101692414
your premise that the AI enthusiast market is a few retards on 4chan is just incorrect man, I don't know what else to say
look at all the people on reddit having a meltie about not being able to run flux
the market is obviously not as big as gamers but it's definitely large enough to be worth servicing using older fabbed parts
>>
File: ComfyUI_00987_.png (1.58 MB, 1024x1024)
1.58 MB
1.58 MB PNG
>>101692630
Thanks- didn't even think about that. Plus I'm not trying to engage people in empty conversation. It's hilarious watching the schizo feel like he's solving that problem
>>
>>101692673
Any card good enough to run Flux well will appeal to businesses. Get a job.
1000 + 1000 is still not enough.
>>
>>101692638
Bullshit. Go to their official site and select it. Last time I checked 1TB it was around 5k
>>
>>101691074
Hands are good now, but I think feet suffer from their total cleaning of most nsfw (nude) images.
>>
>>101692401
how do you prompt for game UIs?
>>
File: file.png (69 KB, 563x1076)
69 KB
69 KB PNG
>>101692694
Where do you think that screenshot is from, retard? $7000 STARTING. FOR THE SHIT MODEL.
>>
I just genned a bunch of dead 2 year old girls floating in a sewer, it's not that hard desu, how fucked are they?
>>
>>101692605
Why not?
>>
File: apple.png (183 KB, 1150x1014)
183 KB
183 KB PNG
>>101692724
?
>>
>>101692691
What's he gonna buy if he "gets a job"? The cards that don't exist? You can't really do any better than 1000 + 1000 right now because the next step up from there is 50 grand. There's no happy medium product segment.
>>
File: FD_00053_.png (1.24 MB, 1024x1024)
1.24 MB
1.24 MB PNG
Model is flawless
>>
>>101692754
>buying the meltdown edition
>>
>>101692758
You can buy one 3090. or if you're feeling frisky you can buy two 3090s. And if you're feeling extra wild you can buy four 3090s.
>>
>>101692765
Concession accepted. Fucking retard
>>
>>101692782
Now I'm confused because I thought dual 3090s is what you meant by "1000 + 1000". That's what I've already got. What on earth were you referring to if not that?
>>
>>101692724
>>101692638
Are you ca or aus? Try a country with normal dollars. It starts @5K
>>
>>101692799
>I'm going to spend $6000 on a computer that will burn out and thermal throttle
>>
>>101692803
Shit you got me there. I'll pivot to another argument now.
>>
>>101692803
Essentially you have NEETs in here that think Nvidia is going to make a ton of money selling 2080s to the NEETs desperate to generate smut at 10s/token which, as a reminder, would be 30 minutes per 200 words.
>>
>>101692782
What's the point of having 2x3090s if imagen models are unable to use their whole available memory and can only use 24GB at a time?

Only use case I see is text gen where it's apparently possible to pool both card.
>>
File: FD_00062_.png (1.05 MB, 1024x1024)
1.05 MB
1.05 MB PNG
>>
File: file.png (1.13 MB, 1024x1024)
1.13 MB
1.13 MB PNG
>>
File: ComfyUI_00995_.png (1.03 MB, 1024x1024)
1.03 MB
1.03 MB PNG
>>
>>101692837
Text gen splits the model per card. Image gen hasn't had that as a requirement (yet) but maybe we'll see it happen here too. But in the case of Flux, the image model fits in 24 GB, the VAE and T5 can go on the other card.
>>
>>101692835
>that think Nvidia is going to make a ton of money selling 2080s
Nice strawman, no one said that. Do another pivot now.
>>
>>101692864
>VAE and T5 can go on the other card
You can do that in comfyUI?
>>
>>101692881
Actually you did, you said they'd sell a ton of 2080s with 64 GB of VRAM. Which isn't true. That card would only appeal to the NEET audience who have more time than cents.
>>
>>101692882
with this patch yeah:
>>101689241

it's great, I have 3090 + 3060 and I've offloaded the vae and text encoders to the 3060, now I can actually run flux in fp16 without my system slowing to a crawl due to hitting 99% vram usage
>>
>>101692882
Yeah there's a custom node someone was using. But it's really just a one liner because you're just doing "cuda:0" for the image model and "cuda:1" for the VAE and T5. So it will be trivial to patch.
>>
File: FD_00071_.png (1.11 MB, 1024x1024)
1.11 MB
1.11 MB PNG
>>101692858
Mooom dads playing with the AI agaaain
>>
>>101692896
I see, reading comprehension / projection. ctrl+f "selling" in this page, only (you) said that word. So again, nice strawman, try following the conversation and not just the words in your head.
>>
>>101692920
>>101692926
Oh that's nice!
>>
>>101692943
yawn, hold your breath anon 64 GB GPUs any day now for $500
>>
File: ComfyUI_00997_.png (1.34 MB, 1024x1024)
1.34 MB
1.34 MB PNG
>>
>>101692965
Nice pivot, again. So you lost on your "can't get a 192 mac under $9500", also lost on the "3090+3090" and now on your last strawman. You can go to bed and let adults converse. thumbs up emoji.
>>
File: ComfyUI_00998_.png (1.86 MB, 1024x1024)
1.86 MB
1.86 MB PNG
>>
>>101692978
That's nice. Can you do one with background decor not just plain color?
>>
File: ComfyUI_Flux_01859_.png (899 KB, 1024x1024)
899 KB
899 KB PNG
>>
>>101692864
>Image gen hasn't had that as a requirement (yet) but maybe we'll see it happen here too
I hope that's the case
>>
>>101692710
I just added "[game] with HUD"
>>
File: ComfyUI_00999_.png (1.48 MB, 1024x1024)
1.48 MB
1.48 MB PNG
>>101693035
>>
File: file.jpg (96 KB, 1280x768)
96 KB
96 KB JPG
>>
File: file.png (730 KB, 1024x1024)
730 KB
730 KB PNG
>>101693008
I don't see how this changes that you're fucked. Good job you won the argument in your brain. Nothing changed. I hope you sleep well tonight still being unable to run Flux.
>>
File: ComfyUI_Flux_01861_.png (936 KB, 1024x1024)
936 KB
936 KB PNG
>>101693040
>>
File: FD_00085_.png (1.48 MB, 1024x1024)
1.48 MB
1.48 MB PNG
>>101693040
Are you foot fags just fully empty balled at the moment or what
>>
File: FLUX__00026_.png (1.01 MB, 1024x1024)
1.01 MB
1.01 MB PNG
what is that example I've seen, the "counter the strike(?)" thing that sold everybody
>>
>>101693055
I like it
>>
>>101693093
>unable to run Flux
Fuckdamnit even my 3090 can't run it now? :/
>>
File: file.png (1.11 MB, 1024x1024)
1.11 MB
1.11 MB PNG
>I'm sure someone out there bought a $9000 Mac for AI, but not me I'm poor
>>
>>101692470
>>prompt leaks as text
Nothing new, affects DallE as well
>>
File: file.png (553 KB, 1024x1024)
553 KB
553 KB PNG
>>101693126
>>
>>101693187
>>
File: 240zbmalo2gd1.png (1.54 MB, 1152x896)
1.54 MB
1.54 MB PNG
>>101693106
>>
>>101693126
Just set it to fp8 in the weight_dtype. You can run it.
>>
File: vnfg6082c3gd1.jpg (99 KB, 1280x1024)
99 KB
99 KB JPG
>>101693251
>>
>>101693230
kek, how can this model be so good at text, dalle3 couldn't do this coherently even without the slur
>>
File: ComfyUI_Flux_01862_.png (951 KB, 1024x1024)
951 KB
951 KB PNG
>>101693097
Better feet so far just by asking for iphone selfies, some inpainting it will get there
>>
File: file.png (1.35 MB, 1024x1024)
1.35 MB
1.35 MB PNG
>>101693230
>>
>>101693264
Yeah with fp8 you can.
You can't run fp16 though even with two 3090s, that NEET person is off their meds.
>>
How to use second 3090 with flux?
>>
File: iszjlv9k29gd1.png (2.04 MB, 1504x768)
2.04 MB
2.04 MB PNG
>>
>>101693302
not him but you can do fp16 on 3090 + one other card if you offload the VAE to the second card
that brings the vram usage down just enough that it becomes usable

could probably do the same with a single 3090 by offloading the vae to CPU, but I imagine that would mean very slow vae decodes
>>
File: file.png (785 KB, 1024x1024)
785 KB
785 KB PNG
>>101693302
>>
File: 5o7aiu9k29gd1.png (1.91 MB, 1504x768)
1.91 MB
1.91 MB PNG
>>101693322
>>
>>101693302
>>101693329
forgot to link post with required comfyui patch: >>101689241
>>
>>101693302
fp16 works fine on my 3090 unless comfyui is gaslighting me and running in fp8 mode or something
>>
File: n9d4vx9k29gd1.png (1.7 MB, 1504x768)
1.7 MB
1.7 MB PNG
>>101693348
>>
>>101693302
I can do FP16 gens on a 4080, it just takes 15 minutes per image
>>
Patiently waiting for a module to handle layer offloading to a different GPU...
>>
File: pn4iv0ssp2gd1.jpg (53 KB, 1152x896)
53 KB
53 KB JPG
>>
>>101693369
that's overflow to RAM
>>
File: ComfyUI_Flux_01868_.png (981 KB, 1024x1024)
981 KB
981 KB PNG
>>101693098
If a model this good can do feet then it needs to be pushed its limits to unlock its capabilities. Don't forget this is what Dalle can do
https://desu-usergeneratedcontent.xyz/g/image/1698/69/1698697621924.jpg
https://desu-usergeneratedcontent.xyz/g/image/1698/66/1698664670985.jpg
https://desu-usergeneratedcontent.xyz/g/image/1696/63/1696638573669.jpg
https://desu-usergeneratedcontent.xyz/g/image/1696/61/1696614127291.jpg
https://desu-usergeneratedcontent.xyz/g/image/1696/63/1696639693066.png
https://desu-usergeneratedcontent.xyz/g/image/1696/62/1696625840562.jpg
https://desu-usergeneratedcontent.xyz/g/image/1696/64/1696642255370.png
https://desu-usergeneratedcontent.xyz/g/image/1696/46/1696466761637.jpg

Don't forget what they took from you
>>
>>101693382
do DiT models actually have layers like a language model?
>>
>>101693362
are you in linux? I think it's windows 3090 people doing the vae offload thing because the OS overhead is a little higher
>>
Anyway, I promised I'd to this, here we go:
gradio web slop interface so you can try out dev/schnell without replicate limits, up to 4 images at once

https://lodging-traditional-working-form.trycloudflare.com/
>>
File: FD_00104_.png (988 KB, 1024x1024)
988 KB
988 KB PNG
I dunno I just don't get the whole feet thing.
It will be an interesting day when DallE finally leaks though
>>
Forgot to mention in >>101693452, the NSFW checker is disabled since replicate allows that when using the api
>>
>>101693410
flux isnt getting close to that unless someone finetunes it for equally as long as the base model was trained. it's simply a massive gap in data quality
>>
>>101693362
I'm in the same boat here. I'm at 23.6/24 vram. Maybe the problem is with people who didn't turn off the aggressive OOM protection thing that was causing problems with LLMs a few months back?
>>
>>101693410
I don't get feet people. What's special about them? Do you also get off on hands? Thats just wierd
>>
File: ComfyUI_Flux_01874_.png (1.1 MB, 1024x1024)
1.1 MB
1.1 MB PNG
>>101693458
>It will be an interesting day when DallE finally leaks though

This model is already Dalle tier
>>
>>101693583
its good but lets not lie to ourselves here.
>>
>>101693469
>NSFW checker is disabled
https://files.catbox.moe/wccd72.jpg
still struggles
>>
File: ComfyUI_Flux_01875_.png (1.07 MB, 1024x1024)
1.07 MB
1.07 MB PNG
>>101693565
Not necessarily. The female body is a work of art.
>>
>>101693624
yeah that's just the model
>>
>>101693362
Do you see your vram drop after each gen and then climb up again? I think it's unloading the model to fit the text encoders and then unloading them to fit the model. So you are running the model fine, but all that swapping slows things down.

I think.
>>
File: ComfyUI_Flux_01876_.png (1.05 MB, 1024x1024)
1.05 MB
1.05 MB PNG
>>101693548
I just tuned my prompt and started getting much better feet. It's all possible with the right engineering. Anything is, we just don't have precise control over style hence the usual shitty gens.
>>
>>101693563
it's not that it doesn't work, it's that the whole system chugs during the vae decode and the decide takes like 30seconds when it should be 2. offloading the vae solves this, makes decoding fast and keeps the system usable during generations by reducing vram usage just a few percent
>>
>>101693309
>>101689241
>>
File: ComfyUI_00321_.png (1.09 MB, 1024x1024)
1.09 MB
1.09 MB PNG
>gonna need to wait one week for A1111 support
>gonna need to wait one month for kohya-ss support

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/16311#issuecomment-2266219185

We won, but at what cost
>>
>>101693689
I can make them drip milk just like Dalle, not sure I can post that on a Christian forum
>>
I'm out of the loop, what's this about flux disappearing?
>>
>>101693825
check the image at >>101688297 and ask yourself if (((they))) will allow this
>>
>>101693831
What was it? Doesn't return anything
>>
>>101693831
that link is dead
>>
>>101693831
Deleted. What was it?
>>
>>101693849
>>101693856
>>101693861
https://desuarchive.org/g/thread/101685374/#101688297
>>
prolly a swastika or some other 14yo edgy shit
>>
Baker
>>
NEW
>>101693167
>>101693167
>>101693167
>>101693167
>>101693167
>>
>>101693913
Thank you baker!
>>
Baker?
>>
>>101693917
>>101693913
kek links to sdg

baker still needed
>>
baker save us from sdg goons
>>
>>101693864
oh thats a dead baby
>>
this bred getting stale... can't breathe .. baker help..
>>
When we needed baker most...
>>
Did baker die? 200 seconds until I can try
>>
i'm baking hol' up
>>
>>101693990
>>101693864

>Didn't share the actual file, just a screenshot of something genning on his screen that could be photoshopped

I'm not buying this one bit
>>
>>101694075
He shared the prompt, try it yourself
https://desuarchive.org/g/thread/101685374/#101688357
>>
No collage cos I'm not the baker

>>101694073
>>101694073
>>101694073
>>
>>101693731

Let Kohya cook.
>>
>>101694086
Tried it, didn't get anything close to that picture, just bad cgi rendering of a skinny dreamshaper woman lying on the ground with some blood on her stomach, her nips are there but it's mostly a strange bulging underwear merged with her genitals. https://files.catbox.moe/jlzn21.png



[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.