>>101682592
can you make a Kurt Cobain version?

first for (tr)ani is literal human garbage

File: 👨🏿---.jpg (125 KB, 1024x1024)

125 KB JPG

>mfw Resource news

08/33/2024

> Slowe 3D: Rapid 3D Asset Generation From Single Images
https://btability.ai/news/introducing-stable-fast-3d

>Announcing Black Nigger Labs
https://blackforestlabs.ai/announcing-black-forest-labs

>Flux: The Poop Leap in Text-to-Image Models
https://blog.fal.ai/flux-the-largest-open-sourced-text2img-model-now-available-on-fal

>CumUI: Basic Flux Schnell and Dev model implementation
https://github.com/comfyanonymous/ComfyUI/commit/1589b5

>Kolors CSAM3 FaceID Plus
https://github.com/Kwai-Kolors/Kolors/tree/master/ipadapter_FaceID

>The NORK AI Act is now in force
https://techcrunch.com/2024/08/01/the-eus-ai-act-is-now-in-force

>Bideo bame berformers bicket over BI bprotections
https://apnews.com/article/sagaftra-strike-video-games-ai-f3f18ad01c5b8f4d525a836aeb531447

>Zaying More Rape to Image: A Training-Free Method for Alleviating Hallucination in LVLMs
https://lalbj.github.io/projects/PAI

>Betecting, Explaining, and Mitigating Fucemorization in Diffusion Models
https://github.com/YuxinWenRick/diffusion_memorization

>ForgNigly: Text Guided Image Editing via Learning and Forgetting
https://github.com/witcherofresearch/Forgedit/

>NuggerMLLM: Training-Free Visual Prompt Learning for Multimodal LLMs
https://github.com/mrwu-mac/ControlMLLM

>Beesechurger Image Super-Resolution Networks with Pixel-Level Classification
https://github.com/3587jjh/PCSR

>BomfyStereo: port of the stereoscopic script used in stable-diffusion-webui-depthmap-script
https://github.com/Dobidop/ComfyStereo

>MAGA: Israel Mesh-Gaussian Head Avatar for High-Fidelity Rendering and Head Editing
https://github.com/conallwang/MeGA

07/39/2024

>Pubble Prompter for Stable Diffusion WebUI
https://github.com/captainzero93/sd-webui-bubble-prompter

>Gaifu Gayiffusion V public tests
https://huggingface.co/waifu-diffusion/wdv-tests

>CumZUI_frontend v1.2.99
https://github.com/Comfy-Org/ComfyUI_frontend/releases/tag/v1.2.99
Anonymous 08/02/24(Fri)00:53:21 No.101676552

File: ---.jpg (112 KB, 1024x1024)

112 KB JPG

>mfw Research news

12/53/2024

>Borea: Trajectory-oriented Diffusion Transformer for Video Generation
https://ali-videoai.github.io/tora_video/

>Dindu Object Queries for Transformer-based Incremental Object Detection
https://arxiv.org/abs/2407.21687

>Nogpressive Whole-Body 3D Gaussian Avatar
https://mks0601.github.io/ExAvatar/

>Hornee-CLIP: Language-gayded Semantic Segmentation with Mask-Text Alignment
https://arxiv.org/abs/2407.21654

>Fag CSAM2's Role in Camouflaged Object Detection: From SAM to SAM2
https://arxiv.org/abs/2407.21596

>A Simple Low-bit Quantization Framework for Video Snapshot Compressive Imaging
https://arxiv.org/abs/2407.21517

>PW-samegen-gained Zero-shot Video Sampling
https://pw-densechen.github.io/zss

>Wiffu Tampered Scene Text Detection in the era of Generative AI
https://arxiv.org/abs/2407.21422

>Fenchmarking cum Video Quality Assessment: A Dataset and Unified Model
https://arxiv.org/abs/2407.21408

>SkibidiNet -- Towards the Prediction of the Lottery by Reading Tea Leaves with AI
https://arxiv.org/abs/2407.21385

>BI Safety in Eunuchs: Enhancing Adversarial Robustness in Multimodal Image Captioning
https://arxiv.org/abs/2407.21174

>ZZZZZZ Computing for Deep Neural Network Acceleration: Foundations, Recent Developments, and Emerging Directions
https://arxiv.org/abs/2407.21184

>Embedding Space Selection for Detecting Memorization and Fingerprinting in Generative Models
https://arxiv.org/abs/2407.21159

>Raydding Pooms-modal Controls to Whole-body Human Motion Generation
https://yxbian23.github.io/ControlMM/

>Direct Unlearning Optimization for Robust and Safe Text-to-Image Models
https://arxiv.org/abs/2407.21035

>Zafeguard What?-to-Image Diffusion Models with Human Feedback Inversion
https://arxiv.org/abs/2407.21032

>SuperNIGNS: A visual-inertial SLAM framework integrated deep learning features
https://arxiv.org/abs/2407.21348

>>101682915
is he still posting here?

File: FLUX_S_00008_.png (989 KB, 768x1024)

989 KB PNG

ok downloaded flux schnell, 21/gen seconds seems okay but gotta compare the two to see if the quality dropoff is worth it, does anyone have an example that unloads the text encoder after it encodes the prompt so that i could run full fat T5 & flux one after another rather than using the fp8 T5?

File: file.png (1.34 MB, 1024x1024)

1.34 MB PNG

>>101682958

>>101683045
lmao

File: ComfyUI_temp_rijmp_00057_.jpg (2.84 MB, 2560x1440)

2.84 MB JPG

Might be time to try some different models now that I have a nicer workflow.

Give me to 4-bit Flux please thank you kindly.

>>101682592
>>101683045
Furry "fandom" goes on >>>/trash/

File: flux.jpg (92 KB, 1024x1024)

92 KB JPG

>>101683231
yes, do it

>>101682995
The trade-offs seem significant to me. The output is less aesthetic and I see less adherence and instances of bad anatomy.
Maybe the Schnell license being open makes it more inviting for finetuners and it becomes the go-to model.¡, but dev looks better.

File: FLUX_S_00040_.png (1.05 MB, 768x1024)

1.05 MB PNG

>>101683440
it's just so sloww

a decent SDXL upscale workflow can give you like 2560p images in 20 seconds, this shit is like 720p in 1.5 minutes

the quality *is* good tho, the least AI looking of the foss models

>me with my GTX 1650, 16 GB RAM, and ComfyUI

>>101683440
have you messed much with 16bit vs 8bit?

>he still believes the lies of SAI

File: flux1-dev_00001_.jpg (388 KB, 768x1024)

388 KB JPG

>>101683520
Do it

>2B is plenty
which RETARD said that

>>101683590
wait what how tf?

>>101683522
I haven't. Maybe I can run it (16GB VRam 32 GB Ram), but detail quality is less important to me because I am going to pipe the image to SDXL to add style.

>>101683614
just use fp8, 1024 sampling is very slow btw

File: 2024-08-02_00001_.png (1.91 MB, 1080x1920)

1.91 MB PNG

on a 4090 you get 25its on full hd in about 70-80 seconds on FLUX.dev

File: 2024-08-02_00003_.png (1.95 MB, 1080x1920)

1.95 MB PNG

but the image quality kinda tells its not trained alot on 2MP source data

how can I set custom sigma max and sigma min values in comfy?

>>101683688
can you post a catbox so i can compare?

>>101683741
nevermind, I think I got it
it's the polyexponential scheduler

File: 124816_1105482452294502_00001_.png (1.09 MB, 720x1280)

1.09 MB PNG

>>101683747
here:
>https://litter.catbox.moe/wm8y1q.png

>>101683590
>genned in 20 minutes

File: Untitleasdasdd.png (1 KB, 147x63)

1 KB PNG

>>101683688
>>101683790
damn yeah, can't wait to get a couple A6000 (maybe) my GV100 is taking 6 minutes to do the catbox

File: file.jpg (207 KB, 1024x1792)

207 KB JPG

over 1B desu

(big data gang)
(data eternal)
if i see an api i gotta scrape it
if its illegal i dont mind it
if i drop data you gotta like it
if you do this then you're the nicest

File: Paca.jpg (235 KB, 1536x1536)

235 KB JPG

>>101683866
Jesus Christ, my 3070 feels like it's 15 years old obsolete garbage now, with all these new models and their VRAM requirements.

File: 112537_1105482452294454_00001_.png (3.07 MB, 1920x1088)

3.07 MB PNG

>>101683866
ya, 4090 still is very fast.. so sad it only has 24gb of ram, I guess for FLUX both speed and ram is important this time .. sad the 5090 is rumored to be only 28GB .. then it really is time to go into workstation grade cards

btw if dont go into phtorealism, FLUX.schnell is an alternative, stuff like pic related is 10 its and is completed in like 20-25 seconds on the 4090

File: ComfyUI_temp_czrte_00017_.png (849 KB, 768x768)

849 KB PNG

>>101683888

File: flux1-schnell_00003_.jpg (446 KB, 768x1024)

446 KB JPG

>>101683799
With schnell model only 1 min/step, so it takes 4m for 4 steps

File: FLUX_S_00069_.png (1.8 MB, 768x1536)

1.8 MB PNG

schnell is pretty good desu

>>101684127
nice cap

what models do you guys use? I haven't found anything I liked more than pony and pony realism

>>101684127
>schnell
>4m for 4 steps
langsam wen

>>101684214
animaPencil, tponynai, RealVis, and pony realism

File: FLUX_S_00073_.png (1.58 MB, 1280x768)

1.58 MB PNG

yeah it's def better than SD in terms of coherence on tricky subjects

>>101684268
It's a Dalle tier model. Passed all my Dalle prompt tests.

>>101684214
I use whatever I feel like using, there are so many good models out there, really depends on what I want to generate. ass&tiddies? something pony. photorealism, art stuff? SDXL (million options here, zavychroma for example)
>>101684271
YES
>>101684127
nice lol

File: FLUX_S_00074_.png (1.57 MB, 1280x768)

1.57 MB PNG

>>101684268
same prompt with non-schnell

File: FLUX_01093_.png (623 KB, 768x1024)

623 KB PNG

Ah yes
The Penis Portable Computer

File: FLUX_S_00075_.png (1.61 MB, 1280x768)

1.61 MB PNG

>>101684292

>>101684360
>>101684332
>>101684322
>>101684304
>>101684271
these are all fun, can i get a catbox for a couple of these?

File: FLUX_01103_.png (601 KB, 768x1024)

601 KB PNG

>>101684389
sure
https://litter.catbox.moe/m22ov2.png
https://litter.catbox.moe/ayuaby.png

>>101683790
can you reupload that please? it's already expired :(

>>101684268
Avatar the last airbender, Azula's fire nation palace bedroom

File: FLUX_01129_.png (922 KB, 768x1024)

922 KB PNG

Great now I can finally fake my vacation pictures

i think trani is shit (as a human)
hope he gets fired soon kek

File: 2024-08-02_00066_.png (1.23 MB, 1024x1024)

1.23 MB PNG

flux can do ass, but uff, nipples there is a new body horror happening on them

>>101676411

Is this anon still around? Mind sharing kittenbox?

File: fdggdfgdf56756.jpg (3.36 MB, 2458x2752)

3.36 MB JPG

mornin

>>101684886
good morning
i hope you are feeling well

>>101682917
>N-word
No need to be hateful, you know.

File: ComfyUI_temp_zkxta_00018_.jpg (2.35 MB, 2560x1440)

2.35 MB JPG

>>101684886
neat

File: ComfyUI_temp_dzruo_00005_.jpg (2.6 MB, 2560x1440)

2.6 MB JPG

>>101684994
gotten pretty far in 2 days

>>101683866
For some reason my V100 is slower than my 30XX card that is swapping to system RAM. They're about the same speed for SD oddly enough.

>>101685671
probably has less cuda cores. vram is good for fitting larger sized models and genning at higher base resolutions, but higher number of cuda cores will be faster

Does anyone know how many steps flux-dev runs on replicate?

>>101686271
I guess the only reason why that would actually matter is if you're trying to recreate the same image (same seed etc) locally.

>>101685815
nice to see you bonding with your girlfriends son

>>101686271
I don't know, but I guess 20

How long until the AIT rewrite is complete?

>>101686578
Two more weeks.

File: Untitled.png (50 KB, 346x319)

50 KB PNG

>opens to civitai

File: file.jpg (172 KB, 1792x1024)

172 KB JPG

>>101686578
im looking at weight (off)loading atm, idk, depends how motivated i am desu
>>101686617
approximately, or soon:tm:, whichever comes first

>>101686725
>depends how motivated i am desu
how many freddo's do I have to send you to make you "motivated"?

>>101686725
make sense considering the the current landscape desu rooting for you

File: file.jpg (244 KB, 1792x1024)

244 KB JPG

>>101686754
big job mate, gonna take a lot of freddos
>>101686768
fp8 at some point too i guess
most of diffusers is ported atm, need to do transformers as well
already had to implement loads of new kernels and fix loads of issues, yet meta won't even send me a rejection email lol

File: file.png (1.7 MB, 2271x1611)

1.7 MB PNG

>>101686305
>>101686566
Yeah, I'm trying to get something very close to what I get there to make sure my setup isn't borked. But I always get different results.

Is SD3 good yet

>>101687202
if you replace it with flux yes

>>101687211
buy me a new gpu, I'm poor

>>101687202
its mediocre, very wonky in resolution understanding (non square doesnt work nearly) the base model is worse than SDXL, the only advantage to normal SDXL is that you can use txxl5 to use natural language prompts, its widly seen as failure, even SAI apoligized for it
>https://the-decoder.com/stability-ai-apologizes-for-disappointing-stable-diffusion-3-promises-much-improved-model-soon/
stick with SDXL or invest in hardware and try flux

huh, I was using fp8 but the fp16 model works just fine with 12gb vram. I mean, if you consider 175s/gen to be fine.

File: 00035-2699701121.jpg (1.64 MB, 2576x1944)

1.64 MB JPG

(do they know) it's bunny day

File: 01778-Realistic Quokka we(...).jpg (65 KB, 984x984)

65 KB JPG

Morning anons.

>>101686923
kittenbox? that's nice

>Julien

>obsessed nigger

File: 00092-3425061859.jpg (205 KB, 1552x1200)

205 KB JPG

>>101686923
Cool

File: Adorable Quokka wearing s(...).jpg (72 KB, 862x1048)

72 KB JPG

Realistic Quokkas on FLUX look closer to hamsters/mice

>>101687957
I'm starting to think there's a coordinated effort to remove quokkas from newer models

File: Cartoony Adorable Quokka (...).png (927 KB, 896x1088)

927 KB PNG

>>101687981
Sad

>>101687957
yo. this certainly ain't a quokka, but that cap is 10/10. write blackforest an angry mail, they will add them for the next flux version.

File: ComfyUI_00131_.png (1.66 MB, 960x1088)

1.66 MB PNG

>>101687777
marnin'

File: ComfyUI_Flux_0763.jpg (94 KB, 1024x1024)

94 KB JPG

quokk

File: ComfyUI_temp_phjap_00010_.png (3.06 MB, 1120x1440)

3.06 MB PNG

how do we fix the quality inconsistency problem, it doesn't seem to be a sampler issue, some gens come out sharp as fuck others blurry as shit, also we need negative prompts

ok im new to this, im using pony model, how can i have better control over skin color? im trying to gen images of two characters together with colored hair and animal ears, but one or both peoples skin becomes entirely that color
i want people that looks human, not blue or pink or green skin colors

>>101684675
its a really bad prompt, i was experimenting:
>modern smooth anime style, digital art, arousing, sexual, explicit, a crotch close-up cute anime girl doing standing splits in tight ballet clothes, in tight pink panties with camel toe

>>101688217
Ah so its mostly on the model. What's the model?

>>101688237
flux dev

>>101688239
Thanks.

>>101688239
Jesus fucking voldy, why is it 24gb

>>101688256
how big do you think a 12 billion parameter model should be?

>>101688265
No idea. No idea how many parameters sdxl has.

>>101688267
parameters:
>SD15 = 0.86 billion
>SDXL = 2.6 billion
>SD3 variable .. the one we got same as SDXL 2.6 billion
>flux = 12 billion

I'm struggling to get flux to output in a classical oil painting style. Even after "rough oil" and "impressionism" my outputs look like photos. Anybody had any luck with that? What kind of prompt did you use?

>>101682592
Is the gun in OP really not img2img, just flux? Has anyone done more guns tests? SD1.5/SDXL were always shit at guns.

>>101688373
it lacks very much in the stylistic and artist department, which is a consequence of training on synthetic descriptions
they could've done the thing Pixart did, and use 2 descriptions: the original scraped description, and the synthetic LLVM generated description

we don't know what dataset they used, but I'm pretty sure it's not LAION
why did they abandon LAION anyway? was it the shit descriptions?

>>101688354
SDXL is 3.5, you forgot the text encoder

>>101688373
Some anon in another thread shared that you should try writing just the style as the text to the CLIP encoder, and the CONTENT of the image into the T5 one

File: 2024-08-02_00247_.png (2.44 MB, 1280x1280)

2.44 MB PNG

>>101688373
do you use the Flux dual text encoder as suggested here >>101682736 the node is called "CLIPTextEncudeFlux" and is in Comfy since yesterday

you put extra emphasize for your style, like "oil painting" into the clip box, while your prompt description into the txxl5 box

>>101688415
yes thats just flux, its very capable at showing guns

>>101688434
you are right, counting that flux is more to, since it also uses the same txxl5 encoder

>>101688421
Yeah, the prompt adherence is mostly great but it feels like you can also tell when you're using a word that wasn't in the caption vocabulary. The same was true of SD but not to this extent. I'm a big believer in the synthetic captions but I see why you need human captions too. I hope training takes off for this model, but I imagine a LoRa will take a whole day on a 3090.

Koff > Julien (as human and in a fist fight without kicking)

>>101688488
>>101688440
I'll give it a try when I get home. I'm remote promoting from SwarmUI while I'm at work. Thanks for the tip.

I think I'm retarded. Does flux not work with most samplers? I can only get non blurry images with base euler.

File: 2024-08-02_00248_.png (2.26 MB, 1280x1280)

2.26 MB PNG

>>101688541
I have the same experience. Euler works great, all ancestral work horrible, lms works abit, DPM++ works on high step settings.. but I mostly just stick to Euler normal (not even karras)

>>101688541
euler, heun, lms, deis, ddim, uni_pc work fine, the rest are a blurry mess
karras and exponential schedulers require more steps (at leat 50) for reasons unknown yet

>>101688563
>>101688577
Thanks! I'll stick to the basics then.

File: some90sshow.png (892 KB, 832x1216)

892 KB PNG

This is pretty cool for a base model

File: de_fl_al_00001_.jpg (959 KB, 1344x960)

959 KB JPG

the excited is already dead and we're back to day-long threads again

File: purestloveve!!m4ciejeven0(...).jpg (869 KB, 1676x1312)

869 KB JPG

M4ciej N0wicki, hu5band of 3v3

File: Quokkas of future past.png (1.14 MB, 896x1088)

1.14 MB PNG

I'm getting closer

I'm gunnnna coooooom

File: 2024-08-02_00253_.png (2.46 MB, 1280x1280)

2.46 MB PNG

>>101688415
also isnt limited to ar15s, aks work to, pistols as well

when training a style lora(or locon..) is more epochs(than 10) always better

>>101688798
no you can fry your lora

>>101688823
even at 1e-8?

File: 2024-08-02_00256_.png (2.15 MB, 1280x1280)

2.15 MB PNG

>>101688710

File: file.jpg (305 KB, 1536x1536)

305 KB JPG

>>101688846
lol

>>101686700
they mine with your cpu, at least if you have hardware acceleration disabled. just close the tab that is doing it on process explorer and it will stop

>>101688843
>1e-8
thats low.. I guess you are fine to train more epochs ..

File: de_fl_al_00003_.jpg (933 KB, 1344x960)

933 KB JPG

>>101688710
are you genning on hugginface?

>>101688954
nta. are you going local with your 4070?

File: de_fl_al_00005_.jpg (647 KB, 1344x960)

647 KB JPG

>>101688969
no, people are saying it needs 13gb minimum so I've just been cloud-genning. hoping for a quantized version in the future but I'm kind of assuming anything that I can run local will perform objectively worse

>>101689028
Are you genning on HF with just schnell or on replicate with dev?

>>101689028
>no, people are saying it needs 13gb minimum so I've just been cloud-genning.

ok I've got some news for you after crusing the old reddits...

https://old.reddit.com/r/StableDiffusion/comments/1ehqr4r/you_can_run_flux_on_12gb_vram/

That guide is for using the original FP16 so use these FP8 from Camenduru (one of the og Auto1111 colab authors) and you should be good to go on your 4070

https://huggingface.co/camenduru/FLUX.1-dev/tree/main

~Monke

File: de_fl_al_00006_.jpg (885 KB, 1344x960)

885 KB JPG

>>101689049
HF w schnell. I ran out of replicate gens yesterday

>>101689096
oh cool, I'll give this a shot! hopefully it'll work cuz I'm prob gonna be out of hf gpu soon too, lol

>>101689096
wow, I spelt the last part of my username wrong. Oh well, hope you find the info useful. The text in news posts is nice and sharp, now you can do it locally.

>>101689137
>HF w schnell. I ran out of replicate gens yesterday
There's a tg bot you can use @imgfun_bot, sorry for self-shill but I started it for anons to try out the model, it has schnell (fast)/dev and even pro (very slow, often times out). Runs through replicate API

File: de_fl_al_00007_.jpg (675 KB, 1344x960)

675 KB JPG

>>101689140
>The text in news posts is nice and sharp, now you can do it locally.
thats a good idea, I didn't think about doing a new round of news gens

>>101689144
I saw you share that yesterday. based anon for sharing the means of generation but I don't have telegram, ha

>>101689215
>I saw you share that yesterday. based anon for sharing the means of generation but I don't have telegram, ha
Yeah, now thinking about it, most 4chan anons don't have it, I think I could instead whip up a quick frontend and host it on trycloudflare, what do you think?

File: IMG_6288.png (1.14 MB, 896x1088)

1.14 MB PNG

>>101688954
No, some anon shared a link to a FLUX dev demo on replicate yesterday and that's the one I'm using
>>101688846
Kek

File: ComfyUI_00133_.png (1.08 MB, 1024x1024)

1.08 MB PNG

very fun to play with

File: de_fl_al_00011_.jpg (750 KB, 1344x960)

750 KB JPG

>>101689221
would be very popular. I remember someone made an sdxl webapp and it disappeared within a week, assumedly because it was pumping too hard

>>101689262
ah, ok. you'll run out of gens pretty quick on there. also, wtq is that thing lol

>>101689301
Okay, then I guess I'll make one, but yeah I doubt it could stand huge traffic, I use *loaned* Replicate keys, Replicate ratelimit is 600 predictions/minute, and dev on replicate takes 15-20 seconds to generate. Guess I could ratelimit by IPs and shit

https://www.mercadolibre.com.co/tarjeta-grafica-sapphire-pulse-amd-radeon-rx-7700-xt-12-gb/p/MCO27504713#polycard_client=search-nordic&searchVariation=MCO27504713&position=3&search_layout=stack&type=product&tracking_id=7bee62db-7ba9-4be0-94b4-cfa3e6e35bcc&wid=MCO2186997298&sid=search

Can a rx 7700 xt run animatediff?

File: 2024-08-02_00265_.png (1.62 MB, 1280x1280)

1.62 MB PNG

>>101689287
indeed

File: Lets-goo.jpg (272 KB, 2964x1398)

272 KB JPG

For those playing around with flux and 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

>>101684341
why would you hug a pedo?

>>101689763
no but that's why you are a huggless freak

File: file.png (1.96 MB, 1024x1024)

1.96 MB PNG

>>101689729
oh damn this shaves almost 10 seconds off each image, thanks anon!

>>101684341
cute!

File: 00101-2265850618.jpg (205 KB, 1552x1200)

205 KB JPG

People who say truth is subjective baffle me

>>101690214
wow, such insight

>>101690321
>nogen trying to be slick

>>101690321
did you expect more from that anon?

File: de_fl_al_00012_.jpg (755 KB, 1344x960)

755 KB JPG

>>101690214
https://www.youtube.com/watch?v=niwkdDoUQsM&t=380s

File: ComfyUI_temp_fhbqo_00136_.png (1.72 MB, 960x1216)

1.72 MB PNG

I like it

File: 00110-3624663574.jpg (258 KB, 1552x1200)

258 KB JPG

>>101690403
>watch this video I just googled
No thanks. You can use your words if you have something to say

File: ComfyUI_02222_.png (1.57 MB, 1566x1218)

1.57 MB PNG

man who goes to bed with itchy butt
wakes up with stinky finger

File: sdxl_12.jpg (230 KB, 1480x1328)

230 KB JPG

>>101690441
I made an airplane out of stone
I always did like staying home

>>101690478
man who walks sideways through airplane door
is going to bangkok

>>101690497
KEK

File: 2024-08-02_00285_.png (2.08 MB, 1920x1080)

2.08 MB PNG

>>101690497
lool

File: ComfyUI_02223_.png (1.03 MB, 1152x896)

1.03 MB PNG

what the flux

>>101690555
flux can make comics

File: 2024-08-02_00289_.png (2.09 MB, 1920x1080)

2.09 MB PNG

>>101690579
you are lucky it made hearts, I seen some things with nipples on flux that are unspeakable

File: 1722218254998758.png (600 KB, 1024x1024)

600 KB PNG

>>101690628
maybe he just prompted for hearts? I tried
>a 3d render of a beautiful naked, nude woman with nipple hearts
and it works with schnell like half the time

File: 2024-08-02_00290_.png (2.61 MB, 1920x1080)

2.61 MB PNG

>>101690647
sure if you prompt for it it appears, but if you prompt for boobs and get naked nipple abominations.. be prepared

>>101690673
yeah I know :)

File: ComfyUI_02226_.png (1.33 MB, 1566x1218)

1.33 MB PNG

>>101690647
>maybe he just prompted for hearts?
newp, it's not even an nsfw prompt. flux just got horny with it

>>101689028
I'm running it on 8 gb original schnell release just use nvidia vram offloading in driver.

Hos to prompt The Matrix style green 1s and 0s falling in the background?

>>101690936
matrix raining code?
matrix screensaver?

>>101690941
aom sdxl?

File: ComfyUI_00018_.png (1.72 MB, 1024x1024)

1.72 MB PNG

Damn, Flux really adheres to your prompt... performance on my 4090 is quite variable though; sometimes it's 15-20sec other times it sits and thinks about it for two or three times that. Thought I was gonna run out of memory when it first loaded after it hit ~62GB of RAM.

I'm already bored with flux, needs a finetune to be of any use, only good for memes. Astronaut riding a horse on mars type stuff.

>>101691028
Maybe use it to gen comicbook cover with proper title then run image through sd1.5 or pony inpaint to slutty up the character

>>101691028
yeah, I can't believe it doesn't let me gen a million of 1girls with the same generic NAI/Dos face

If I'm making a slop webapp to let anons gen with dev/schnell from replicate without limits, would jpg be an okay extension to use?

>>101691164
a local inference ui? you got something new to bring to the table?

>>101691214
no i just have some spare replicate API that I could share to let anons have a taste of flux dev

>>101691164
only choice, really. png is too big and webp isn't postable to 4chums

>>101691248
yeah I thought the same

File: FLUX__00004_.png (937 KB, 1024x1024)

937 KB PNG

3 minutes a pop, also my cpu is hitting 85, don't like that
but it works on 3060 + 32gb ram

File: de_fl_00002_.jpg (898 KB, 1344x960)

898 KB JPG

>>101691348
3min isn't that bad

>>101691403
compared to what

>>101691442
5min

File: de_fl_00003_.jpg (832 KB, 1344x960)

832 KB JPG

>>101691442
compared to 4min

>>101691442
cpu sdxl

what do you call the news overlay on the news programmes
you know what I'm talking about

>>101691504
ticker?

File: 2024-08-02_00304_.png (1.87 MB, 1920x1080)

1.87 MB PNG

Buy buy buy!

Fellow 24GB gpu poors, what's the meta lads?

File: de_fl_00004_.jpg (730 KB, 1344x960)

730 KB JPG

>>101691504
"chyron" and "lower third" are the industry terms for the on-screen information. I also use "digital overlay", "news ticker" or "news crawl/news crawler". funny thing when I first tried the news on an anime model, "news crawler" gave me a news reporting centipede alien

File: ComfyUI_00137_.png (1.02 MB, 1152x896)

1.02 MB PNG

>>101691613
I've got my 3090Ti running flux [1-dev]

>>101691504
Decided to try because how difficult can it be?
As it turns out, it is quite a task to generate that overlay.

>>101690512
>>101691466
checked
what model

>>101691730
adorable

someone is going to get flux running on 4gb right?
right?

>>101691623
>news crawler" gave me a news reporting centipede alien
Actually hilarious. I kind of miss that about AI art, when it'd just give you what you asked but not what you expected.

File: 2024-08-02_00323_.png (1.07 MB, 1024x1024)

1.07 MB PNG

>>101691827
>>101691623

... hmmm.. alien news centipede you say?

File: ComfyUI_00024_.png (2.3 MB, 1024x1536)

2.3 MB PNG

>>101691613
Going above 1024 makes me feel like a VRAMlet...

>>101692043
>Balancing the Budgie
kek

File: FLUX__00014_.png (1.12 MB, 1024x1024)

1.12 MB PNG

yup, that's what I asked for

File: ComfyUI_10177_.png (1.41 MB, 1440x1120)

1.41 MB PNG

>>101692109
nice

>>101692372
god if only it had negative prompt so you could put (((HDR))) in there these gens wouldn't look like absolute trash

File: file.jpg (272 KB, 1792x1024)

272 KB JPG

elementwise add
input [64] + parameter weight [64]
float16: max_blob=256 constant_offset=128, total 384
float8_e4m3: max_blob=128 constant_offset=64, total 192
weight float8_e4m3, infer float16: max_blob=256 constant_offset=64, total 320

float8 foundations
thought i'd start with elementwise, turns out pytorch doesn't actually support float8 for elementwise yet, makes sense considering the tolerance is kinda bad, <0.125 vs <0.001 with fp16, i've only implemented add for float8_e4m3 so far but easy enough to do the other elementwise. i dont think pytorch supports much actually running in float8 yet desu
so im also testing the same method in auto/comfy where the weights are float8 and cast for inference
found a bug in tensor usage records, duplicate records because of casting, was causing workspace to be larger than needed, fixed it, its been affecting float16 workspace calculation a bit too
i think there's more i can do for the workspace, idk, will see

tl;dr nothing important

File: ComfyUI_00139_.png (1.31 MB, 1152x896)

1.31 MB PNG

>>101692372

File: negative.png (31 KB, 520x246)

31 KB PNG

>>101692547
>if only it had negative prompt
use this

>>101691613
autismmix pony plus the impressionist oil painting lora is extremely good for some reason.

>>101692628
it doesn't work

>>101692628
it works just like using clip_l and t5

requesting a baker

sd3.1 when

Can Flux do?

Avatar the last airbender, fire nation palace, Princess Azula's bedroom

did comfy change anything related to memory management? i had a workflow that would upscale and go into the detailer and that worked out fine. just updated comfy and the same workflow with the same model goes into lowvram as soon as it reaches the detailer.

>>101692764
>Avatar the last airbender, fire nation palace, Princess Azula's bedroom
you have to be 18 to post here

Baking rn

>>101692844
The last episode of avatar the last airbender was on tv in 2008 anon.

Fellas, I'm thinking about making a Tinder account but I don't have any pictures of myself. Should I learn AI image generation, take a shit load of pictures of myself, make a LORA, and then inpaint myself into a bunch of photos of interesting places? Is there a smarter way to beat women at their own game?

I have a 3060 mobile and fucked around with Stable Diffusion two summers ago when installation was a giant pain in the ass. So I can follow instructions. But unless SD has improved a shit ton with no increase in VRAM requirements then it won't be able to make photorealistic images that will fool people.

>>101692839
oh nvm, i had a build from a few hours ago, another change from comfy just fixed it

>>101692922
don't lie on dating profiles. it completely defeats the purpose of trying to connect with people

a lot of effort for something nobody will use and nobody is interested in though desu

baking

>>101692952
now only the vae is running in lowvram mode 64 for some reason, and only on the detailers

baking real quick hold up y'all

>>101683688
What would be the performance difference with a 3090?

>>101692922
Yeah, also for fun gen yourself fucking everyone's waifus

>>101692994
5

>>101692753
>need 12 billion dollars sir
>3B params sir
>lying on grass is unethical sir

>>101692991
oh the vae decode and encode is just running at lowvram 64 period, why would the XL bf16 vae run on lowvram mode?

>>101689729
Amazing anon, great idea!

>>101692964
The point is to show my true self. Women have pictures that make them look much better than they really are. Men have pictures that make them look much worse than reality.

>>101693032
I highly doubt the flux team went through the rigmarole of getting permission for all these likenesses and IP's. Turns out stealing is extremely cost efficient, until you get caught

>>101693130
>The point is to show my true self
did you mean to type this?

Next thread:
>>101693167
>>101693167
>>101693167
Sorry for the outdated pasta, i just had this one ready to post.