File: FDG_News_00001_.jpg (109 KB, 1216x832)

109 KB JPG

>mfw Resource news

08/01/2024

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

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

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

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

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

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

>Video game performers picket over AI protections
https://apnews.com/article/sagaftra-strike-video-games-ai-f3f18ad01c5b8f4d525a836aeb531447

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

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

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

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

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

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

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

07/31/2024

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

>Waifu Diffusion V public tests
https://huggingface.co/waifu-diffusion/wdv-tests

>ComfyUI_frontend v1.2.7
https://github.com/Comfy-Org/ComfyUI_frontend/releases/tag/v1.2.7

>mfw Research news

08/01/2024

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

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

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

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

>Evaluating SAM2'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

>Fine-gained Zero-shot Video Sampling
https://densechen.github.io/zss

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

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

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

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

>Optical 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

>Adding Multi-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

>Safeguard Text-to-Image Diffusion Models with Human Feedback Inversion
https://arxiv.org/abs/2407.21032

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

>>101673399
>>Discord
>6wUwtcJsr2
What is?

File: FLUX_00242_.png (1014 KB, 768x1088)

1014 KB PNG

>A black and white photobooth film photostrip of a well-dressed pepe frog wearing a Pez and sunglasses. The same well-dressed has different expressions in each photo
From midjourney's showcase page

File: de_fl_00040_.jpg (270 KB, 1344x768)

270 KB JPG

this is my favorite flux so far

>>101673566
prompt pretty please?

File: goo_00040_.png (1014 KB, 1024x1024)

1014 KB PNG

for one this will be an insane instant local meme machine with total freedom

File: FLUX_00264_.png (734 KB, 768x1024)

734 KB PNG

>Give me your wallet

Someone try this on Flux?:
one beautiful woman holds a big red apple in her right hand, one ugly old man holds a small blue apple in his left hand

File: FLUX_00272_.png (841 KB, 768x1024)

841 KB PNG

>>101673586
>You heard her.
>Wallet! Now!

File: 1700288194312997.png (1.59 MB, 1024x1024)

1.59 MB PNG

>>101673604
https://replicate.com/black-forest-labs/flux-dev try yourself
From what I heard the prompt perspective is from the viewer, so you should replace the hand positions.

File: goo_00042_.png (1.14 MB, 1024x1024)

1.14 MB PNG

>>101673600
yea thats what I am looking for? what did you prompt?

File: FLUX_00275_.png (985 KB, 768x1024)

985 KB PNG

>>101673604

File: goo_00043_.png (1.37 MB, 1024x1024)

1.37 MB PNG

>>101673604
>one beautiful woman holds a big red apple in her right hand, one ugly old man holds a small blue apple in his left hand

File: 1715141419936425.png (1.38 MB, 1216x832)

1.38 MB PNG

Flux dev bf16 vs fp8 comparison
https://imgsli.com/MjgzNzEz

>>101673604
picrel

File: ComfyUI_15142_.png (1.41 MB, 1024x1024)

1.41 MB PNG

>>101673618
i went back to the roots of sd prompting - the spam technique (enhanced by chatgpt).
not getting consistent results, maybe 1/5 are passable as transparent-ish

semi-transparent, translucent, slime, see-through, anime style, ethereal, transparent, 

semi-transparent, translucent, slime, see-through, anime style, ethereal, transparent, 
semi-transparent, translucent, slime, see-through, anime style, ethereal, transparent, 
semi-transparent, translucent, slime, see-through, anime style, ethereal, transparent, 
transparent, transparent, transparent, transparent, transparent, transparent, transparent, transparent, transparent, transparent, transparent, 
cute anime-style image of a girl hugging a transparent, jelly-like slime. The girl should have a joyful and affectionate expression as she embraces the semi-transparent slime, which should appear gooey and slightly shimmering. The slime’s transparent quality should be evident, with hints of light reflecting through it. The scene should be colorful and whimsical, highlighting the girl’s warmth and the slime’s playful, see-through nature. Outside, outdoors, forest.
transparent, transparent, transparent, transparent, transparent, transparent, transparent, transparent, transparent, transparent, transparent, 
Tags: cute anime girl, hugging, transparent slime, jelly-like, gooey, joyful expression, shimmering, whimsical

semi-transparent, translucent, see-through, anime style, trees visible through, ethereal, transparent, slime, cute, kawaii, chibi, kawai

>>101673604

File: what.png (1.9 MB, 3095x2855)

1.9 MB PNG

I'm activating the preview thanks to comfyUI manager but it doesn't work, why?

File: FLUX_00294_.png (944 KB, 768x1024)

944 KB PNG

>Bowser said he want his money

>>101673663
anooooon, please share prompts :(

File: goo_00046_.png (1.19 MB, 1024x1024)

1.19 MB PNG

>>101673642
rofl.. okay I guess I have to wait for fine tunes for my slimes, everything else is great tho

also pepe gets interpreted as kermit .. kek

>>101673686
they used nintendo characters in the data? call the lawyers

So are we back to gatekeeping good prompts?

>>101673687
mint pussy is notoriously autistic and gatekeepy

>>101673714
Have you forgotten how to describe an image you see with words?

>>101673708
>>101673686
nice, this is DALLE-3 tier, i remember those memes images

>>101673687
>>101673714
bro you're not even posting gens

>>101673714
>>101673725
just feed it into a llm vision model and you should get an idea, if you dont have any imagiation, even chatgpt could caption an image like that, plus we're all using the same model for now

File: ComfyUI_169840_.png (1.29 MB, 1024x1024)

1.29 MB PNG

>>101673668
I implemented it a few minutes ago so you'll have to update again.

>>101673646
nice

>>101673668
because that is for KSampler node, not Advanced Sampler nodes
do what I said here
>>101673286

Is there any point in getting a second 3090 when using comfyui?
Can it take advantage of that for stuff like FLUX?

>>101673804
can you give me a picture with all the good nodes on it? I don't understand this spaggheti shit :(

>>101673780
Comfy, when are you gonna implement multigpu? I have a 3090 but flux is almost reaching its limit, I have another card that could be used

File: FLUX_00331_.png (679 KB, 768x1024)

679 KB PNG

>They said I have to eat this green glob if I want to be president. That's what they said.
>And I said "Fine". I said "Fine, I'll eat it". That's what I said.
>Here I am. I am eating the green glob. This beautiful green glob. I always like green globs. Love them. Beautiful, beautiful green globs.

>>101673839
[spoiler]schizophrenic[/spoiler]

File: replicate-prediction-z9kw(...).png (1.07 MB, 768x1344)

1.07 MB PNG

>>101673286
That prompt

>>101673780
can I run it on a 3060 with 12gb?

>>101673850
staple prompt for so pony *cough *cough

File: ssssss.png (2.15 MB, 1335x1335)

2.15 MB PNG

>>101673616
>>101673622
>>101673634
>>101673636
>>101673642
Thanks guys

>flux-dev
mine

File: FLUX_00334_.png (846 KB, 768x1024)

846 KB PNG

>Hello, Mr Alien. Or should I say "Shalom!"

>>101673880
kek

>>101673876
no, and it would be dogshit even if you could

File: de_fl_00041_.jpg (268 KB, 1344x768)

268 KB JPG

>>101673839
he's got the green glob eaters vote on lock

File: FUCK.jpg (186 KB, 2629x936)

186 KB JPG

>>101673780
Do you know why your software always unload and reload models? I have enough ram and vram to keep the model in memory, that's slow

What is recommended for comfyui?
Linux or windows? Or doesn't matter?

File: kek.jpg (233 KB, 2545x1750)

233 KB JPG

>>101673780
thanks

File: pepe smile.jpg (11 KB, 244x248)

11 KB JPG

>>101673916
>leave home for a couple weeks
>the next fucking day a Dalle3-tier model gets released
ahgahghaghahghagha

File: 1722507945563498.png (1.49 MB, 1488x840)

1.49 MB PNG

>>101673880
>>101673913
How goes the campaign for captain orange or black lady not a black lady?

File: ComfyUI_09901_.png (567 KB, 1024x1024)

567 KB PNG

behold! my first flux gen

File: ComfyUI_Flux_0147.jpg (161 KB, 1024x1024)

161 KB JPG

>>101673921
linux is seen as better since its less bloaty. but you'll be fine using whatever you're most comfortable with

>>101673953
kek roger rabbit ass pic

>>101673780
I have updated but TAESD still doesn't seem to work
>Warning: TAESD previews enabled, but could not find models/vae_approx/None

File: FLUX_00354_.png (596 KB, 768x1024)

596 KB PNG

>>101673953
lmao

File: ComfyUI_00022_.png (1.11 MB, 1024x1024)

1.11 MB PNG

>>101673931
Ikr, never expected such a day, I'm so fucking happy, SAI BFTOW

File: goo_00062_.png (947 KB, 1360x800)

947 KB PNG

>>101673962
I think it wants Euler different sampler?

also flux doesnt want to make oily or shiny skin

>>101673975
did you downloaded the taesd models and placed them in the folder as instructed? https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#how-to-show-high-quality-previews

File: de_fl_00043_.jpg (251 KB, 1344x768)

251 KB JPG

>>101673947
sorry, the jannies don't like when I talk politics

>>101673953
kek

>>101673962
we've truly enter a new era of generative AI

>>101673969
this might upset the jannies

>>101673990
>look at this big beautiful clone

>>101673969
Thanks anon, I'll probably go with ubuntu then.

File: 1705702489689379.png (976 KB, 896x1152)

976 KB PNG

schnell is kinda weird for me with this prompt, dev doesn't have that issue

>>101673825
here you go
https://litter.catbox.moe/4xgqj6.png

>>101673850
you got better skin exposure kek

>>101673953
kek

Do you have comparison between the three flux models for the same prompt/style?

File: mm.jpg (94 KB, 1278x668)

94 KB JPG

>>101674007
they were already there, but it says that this preview only works with SD1.5 and SDXL?
>The default installation includes a fast latent preview method that's low-resolution. To enable higher-quality previews with TAESD, download the taesd_decoder.pth (for SD1.x and SD2.x) and taesdxl_decoder.pth (for SDXL) models and place them in the models/vae_approx folder. Once they're installed, restart ComfyUI to enable high-quality previews.

Is it possible to set Flux up with WebUI, or will I be forced to move on to ComfyUI?

File: out-0.png (1.08 MB, 1216x832)

1.08 MB PNG

first model that actually creates minecraft gameplay screenshots, this is fun to generate
>>101673962
My first one was a city made of cheese on top of a marble countertop, it's my test prompt for every model

File: sdxl_realism_apple.jpg (541 KB, 2048x1024)

541 KB JPG

>>101673604
same prompt on sdxl + epicrealismXL

File: goo_00067_.png (894 KB, 1360x800)

894 KB PNG

>>101674038
blurry .. i had that to when I didnt use Euler with flux.

>>101674029
>no negative prompt
does flux work with negative prompt?

>>101674065
>first model that actually creates minecraft gameplay screenshots, this is fun to generate
DALL-E 3 could do that too, but yeah, Flux is better

File: ComfyUI_09904_.png (935 KB, 1024x1024)

935 KB PNG

>>101674005
it was too much cfg

>>101674050
>Is it possible to set Flux up with WebUI, or will I be forced to move on to ComfyUI?
I fucking hope it'll be implemented on A1111, first time in my life I run ComfyUI and I already hate it kek

>>101674046
people can barely run one flux and you want them running three

>>101674080
It only outputs cheap minecraft imitations, it gets that it's about blocks but that's it. This one actually tries to put the inventory bar, correct textures...

>>101674093
If you can run one, you can run the other.
And the third one is API only, probably best behind paywall as always.

File: FLUX_00380_.png (962 KB, 768x1024)

962 KB PNG

>Listen here, Jack! You gotta start smoking more weed.

File: ComfyUI_09905_.png (1.72 MB, 1280x1280)

1.72 MB PNG

>>101674069
its not a sampler issue, im using another guider and you can set the cfg range there, this is using 1.0 - 1.0

>there are people ITT who actually give a fuck about models, and will attack/defend other models

why? does someone pay you for this shit?

>>101674070
schnell no
dev might

File: 1699955378408494.png (1.2 MB, 1024x1024)

1.2 MB PNG

>>101674100
Flux is better, but maybe you haven't seen DALL-E 3 actually generate with a proper jailbreak on the API with natural style so it doesn't rewrite the prompt? Here's how it looks

>>101674138
if dev can't do negative prompt then it's DOA

File: goo_00069_.png (1021 KB, 1024x1024)

1021 KB PNG

>>101674082
oky, well I am noodlenoob so idk much about comfy

>>101674068
ya but sdxl uses CLIP as text encoder right? no wonder

File: FLUX_00390_.png (896 KB, 768x1024)

896 KB PNG

>Like I always say, "Flame On!"

Just as a reminder - schenll/dev both can generate ch*ld nudity easily, please do not use them for harmful purposes, I reported both models to my government already.

>>101674141
>you haven't seen DALL-E 3 actually generate with a proper jailbreak on the API with natural style so it doesn't rewrite the prompt?
yeah, you are right, I forgot about that

>>101674171
ESL anon? How are you?

File: ComfyUI_169849_.png (1.19 MB, 1024x1024)

1.19 MB PNG

>>101673975
There's no TAESD for flux yet.

>>101674050
Today is the end of "stable diffusion" and "stable diffusion webui".

File: FLUX_00392_.png (903 KB, 768x1024)

903 KB PNG

>>101674171
>Thank you, Patriot. I will fact check this information, and I will not sleep until I am done.

>>101674171
then you better 99% of realistic models on civitai to ..

>>101674188
>Today is the end of "stable diffusion" and "stable diffusion webui".
Stop coping, comfyfucker

File: ComfyUI_00024_.png (1.03 MB, 1024x1024)

1.03 MB PNG

>>101674188
>Today is the end of "stable diffusion" and "stable diffusion webui".
facts

day1 flux support. auto could never

File: de_fl_00047_.jpg (144 KB, 640x1536)

144 KB JPG

>>101674050
>Is it possible to set Flux up with WebUI
it'll come around eventually but you'll be waiting a while. if you wanna play with the newest toys, comfyui is the only way and has been for a long time now

>>101674115
it failed the biden test

>>101674141
stop showing us how good dalle is. it makes me angry

>>101674188
>Today is the end of "stable diffusion"
tell that to my GPU that can't run flux

>>101674171
apparently, and please don't quote me on this, but you can actually generate ch*ld nudity in your head if you think about it. Best start turning yourself in, anon

>debo fighting tooth and nail for a single (you)

File: 1715267671649210.png (968 KB, 896x1152)

968 KB PNG

>>101674152
just tried, dev can do negative prompt
I put sunlight in neg for picrel
It made the lighting very flat so looks like it can do semantic negatives

>>101674248
comfy biden

>>101674188
weak bait. forge > comfyui

File: FLUX_00408_.png (745 KB, 768x1024)

745 KB PNG

>>101674264

>>101674224
>tell that to my GPU that can't run flux
even on fp8? that asks for 13gb of vram

File: goo_00076_.png (1.13 MB, 1024x1024)

1.13 MB PNG

>>101674224
>it failed the biden test
it fails alot of my tests atm .. but I don't follow the news, what is flux license? is it allowed to be finetuned? if so its a new era
>>101674224
>tell that to my GPU that can't run flux
yea, wont be my goto model for a long time, I dont like to be stuck on 1024x1024

What's the loss of details between fp16 and fp8?

>>101674262
>just tried, dev can do negative prompt
let's gooo

File: ComfyUI_Flux_0185.jpg (98 KB, 1216x832)

98 KB JPG

>>101674050

File: out-2.jpg (142 KB, 640x1536)

142 KB JPG

this thing is wild

>>101674320
thanks dave

File: trump_vs_blue_alien_night(...).jpg (627 KB, 1024x1024)

627 KB JPG

They contracted something

>>101674355
>contracted

>>101674355
uranus

imagine running FLUX on 4 GB VRAM. Yeah, me neither. SDXL ftw.

File: goo_00078_.png (934 KB, 1024x1024)

934 KB PNG

>>101674321
ya 12b parameters has so much potential .. just fill them with the right data

File: de_fl_00048_.jpg (268 KB, 768x1344)

268 KB JPG

>>101674233
this counts

>>101674272
lol this looks like an 80s movie poster
>Honey, I Sniffed the Kids

>>101674277
tfw 12gb 4070

>>101674283
>it fails alot of my tests atm
yea, I'm having a rough time getting anthropomorphic athletes reliably. what are you using for cfg?

>>101674355
>listen glorbgrub, I think your people are great and very hard workers but you have to come here legally

File: FLUX_00430_.png (603 KB, 768x1024)

603 KB PNG

>Hello, chat! I'm back on twitch.
>Today, we're going to play Minecraft.
>My son, Baron, set up a server for us.

So how do i run this on a 3090ti?

File: laugh-point.png (106 KB, 498x498)

106 KB PNG

>>101674271
Sure

File: FLUX_00434_.png (613 KB, 768x1024)

613 KB PNG

>Joe, did you take the diamonds?
>Those were my last diamonds that I needed for my armor.

File: 1721201451276647.png (1003 KB, 1296x920)

1003 KB PNG

Can Flux gen a Laura Kinney? if not lol

File: goo_00079_.png (1.34 MB, 1024x1024)

1.34 MB PNG

>>101674412
>for cfg?
kek I have to change the node, I just used the example workflow, my brain is noodled to a1111 gotta change it to advanced Guider right?

also hands are a problem it seems even on such a model

File: FLUX_00444_.png (829 KB, 768x1024)

829 KB PNG

>>101674438
>What?

File: ComfyUI_15159_.png (911 KB, 1024x1024)

911 KB PNG

>>101674441

File: out-0.png (1.02 MB, 1024x1024)

1.02 MB PNG

jej

>>101674469
Looks like Dalle

File: 1722362277304232.png (974 KB, 896x1152)

974 KB PNG

>>101674289
for DiT, not too significant
for TE, subtle composition issues

>>101674441
yes

>>101674478
kek

>>101674469
>>101674491
Very fast though

>>101674491
ok thanks

File: file.jpg (351 KB, 1792x1024)

351 KB JPG

don't know why i was expecting diffusers documentation to be correct, it never is

tl;dr flux?

File: donald trump is watching (...).jpg (415 KB, 1024x1024)

415 KB JPG

flux
donald trump is watching cute anime

>>101674601
looks likes cute and funny anime and he has some under desk action

>>101674593
>any substance introduced in the smelting of ores to promote fluidity and to remove objectionable impurities in the form of slag

>>101674593
former SAI researchers that were either fired or left the company have started their own, and have delivered an open-source model that is on par with DALL-E 3

File: goo_00088_.png (462 KB, 1024x1024)

462 KB PNG

>>101674412
>yea, I'm having a rough time getting anthropomorphic athletes reliably. what are you using for cfg?
I replaced the basic guider with the cfg guider everything is blurry now .. what cfg to use for flux?

>>101674486
>>101674491
>>101674508
>>101674541

1girls looks juggernatty sdxl as fuck, they probably used a synthetic image set, thats a shame

File: file.jpg (489 KB, 1792x1024)

489 KB JPG

of course they merged the pr with incorrect docs
>hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
that's the wrong shape >:(
shape of img_ids and txt_ids undocumented too

FluxTransformer2DModel
INFO <aitemplate.testing.detect_target> Set target to CUDA
AIT output `Y` shape [[1, 1], [1024, 4096], [64]]
INFO <aitemplate.compiler.compiler> Start to compile AIT model. test_dir='./tmp\\flux'
INFO <aitemplate.backend.target> Loading profile cache from: C:\Users\user\.aitemplate\cuda.db
INFO <aitemplate.backend.profiler_cache> table_name='cuda_gemm_3' exists in the db
INFO <aitemplate.backend.profiler_cache> table_name='cuda_conv_3' exists in the db
INFO <aitemplate.backend.profiler_cache> table_name='cuda_conv3d_3' exists in the db
INFO <aitemplate.compiler.compiler> optimized graph elapsed time: 0:02:05.216291
INFO <aitemplate.compiler.transform.refine_graph> reduced unique ops from 3490 to 794

>>101674636
robin drove the company into bankruptcy ordering too many nodes and fucked up the model then he left

File: goo_00089_.png (1.08 MB, 1024x1024)

1.08 MB PNG

well lol seems its cfg 1

File: sdxl trump is watching anime.png (1.27 MB, 1024x1024)

1.27 MB PNG

at least he(anime) is watching him

File: 1704297415348379.png (1.12 MB, 896x1152)

1.12 MB PNG

>>101674593

>>101674633
kek

>>101674647
CFG works only in flux dev
start small, 1.1 - 1.5

File: de_fl_00049_.jpg (282 KB, 768x1344)

282 KB JPG

>>101674444
>cat girls still have human ears
its over. I'm canceling flux

>>101674508
>>101674541
>wanna flux?
missed opportunity

>>101674647
I'm just hovering around 3.5. I think it started frying over 5

File: humanoid.png (1.16 MB, 768x1024)

1.16 MB PNG

>>101674412
idk, I can get them alright
>anthropomorphic wolf running in the olympics

File: flux donald trump is watc(...).jpg (339 KB, 1024x1024)

339 KB JPG

donald trump is watching tv, anime is displaed on the tv

File: 1708057059028860.png (1.25 MB, 896x1152)

1.25 MB PNG

>>101674705
this gave me an idea

File: photorealistic donald tru(...).jpg (446 KB, 1024x1024)

446 KB JPG

photorealistic donald trump is watching tv, cute anime is displaed on the tv

File: replicate-prediction-rey2(...).png (1.15 MB, 1024x1024)

1.15 MB PNG

>>101674436

>>101674694
kek

File: ComfyUI_Flux_0215.jpg (142 KB, 768x1344)

142 KB JPG

it got the olympics logo right damn

>>101674799
LMAO

File: de_fl_00052_.jpg (265 KB, 768x1344)

265 KB JPG

>anthropomorphic hamster competing in olympic shooting competition
ok thats just a hamster with a gun

>>101674705
pretty goog. it comes and goes for me

File: ComfyUI_Flux_0221.jpg (139 KB, 832x1216)

139 KB JPG

>>101674854
gave that idea to llama3.1 and got this prompt:
>Action-packed shot of a miniature athlete, a charismatic hamster dressed in tiny Olympic attire, proudly holding a toy rifle, standing confidently on hind legs amidst a sprawling, intricately designed shooting range backdrop, with blurred motion and dynamic lighting capturing the intensity of the competition, set against a warm, golden late afternoon sun.
>picrel

is 2x24GB VRAM worthwhile to have or is multi-gpu not a thing yet for the new model

>>101674947
not a thing yet, it fits nicely on a quadro 8000 though

File: de_fl_00053_.jpg (268 KB, 1344x768)

268 KB JPG

>>101674928
strikingly similar

File: de_fl_00056_.jpg (273 KB, 1344x768)

273 KB JPG

he doesn't look like a competitor. he just looks like he's gonna start shooting into the stands

File: Turbo.jpg (348 KB, 1536x1536)

348 KB JPG

Man, this new model has breathed a lot of new life into this thread and that's nice.

File: 00128-TFT_12403052.png (685 KB, 768x1280)

685 KB PNG

cat

File: de_fl_00059_.jpg (292 KB, 1344x768)

292 KB JPG

>>101675285
don't get used to it. sigma, hunyeon, and even aura to a lesser extent all created brief surges in posting but it died off pretty quick too

File: file.jpg (547 KB, 1792x1024)

547 KB JPG

>max_blob=7767720000
not sure if this is too much, might need to fix the memory planning/investigate further. already found a ~5% savings by disabling constant folding
everyone's running it in fp8 though but someone tell me what's the total vram usage anyway, weights would be like 11gb in fp8, i think i saw 17gb mentioned, so 6gb workspace, idk, maybe ~8gb workspace in fp16 isn't so bad
i should implement fp8

does negative prompt work on flux? Comfy says that this architecture is only supposed to work on CFG = 1 (no negative prompt)
>>101675293
>It's a CFG issue, these models don't work with CFG. Set it to 1.0 if using the regular sampler node.
>>101675408
>what? if this model doesn't work with cfg, that means it doesn't work with negative prompt?
>>101675441
>Exactly. Positive prompts only.

>>101675386
it's way better than any local models we got so far, it won't die, it will kill the others

>>101675506
>no negative prompts
I don't get it. Isn't the model ultimately doing some variation of "predict the noise" and then the sampler gradually denoises the image? Why wouldn't CFG / negative prompts work?

>>101675506
works with dev

File: de_fl_00070_.jpg (560 KB, 1344x768)

560 KB JPG

>>101675523
>it will kill the others
if it can fit on my cards in the future, it can stick. 'better' doesn't count for much when almost no one can use the model

File: neotrump14.jpg (490 KB, 1344x768)

490 KB JPG

>Error
>You have reached the free time limit.
final output

>>101675606
Other people's inability to use a car doesn't mean I can't use my car.

File: de_fl_00071_.jpg (795 KB, 1344x768)

795 KB JPG

>>101675627
ok

>>101675662
Thread schizo

File: 00204-1149992481.jpg (1.57 MB, 2576x1944)

1.57 MB JPG

1.5 still better

>>101675523
>it will kill the others
Not a model that won't run on my machine. Yes, I'm a vramlet.

File: ComfyUI_169886_.png (1.3 MB, 1024x1024)

1.3 MB PNG

>>101675506
It's not the architecture it's because flux-dev is a guidance distilled model.

>>101675625
Hopefully this doesn't count as self-promotion, but so that you and other anons can try, here's a temp Telegram (I just use it a lot myself) bot with DALL-E 3/Flux Schnell (fast)/Flux dev:

@imgfun_bot

Just do /start for it to output usage, then just use it. It uses replicate with billing enabled, you can request up to 4 images at once, schnell is around ~5 seconds total, dev can take up to 40 seconds for all images sadly.

I'll turn it off later, there are no ads, nothing of the sort. Just so you can try it.

988m total

File: dave.jpg (132 KB, 1024x1024)

132 KB JPG

>>101675769

>>101675861
Yeah I know it might look like that, but I just want anons to taste the future, and so you can compare it to how bad dall-e 3 is

File: 0.jpg (189 KB, 1024x1360)

189 KB JPG

JYVSW2

Neeeeeeeerrrrrrrrrrddddddddsssssssssssss
(good morning edition)

does this work with a 3060 at acceptable speeds?

>>101676070
lol

File: 430x.jpg (478 KB, 1792x1024)

478 KB JPG

>>101675870
I'm not convinced DALLE3 automatically becomes bad if a new good model comes out, they both have their strengths

>>101676132
Well yeah, but still, dalle3 has too many restrictions

>>101676074
:(

File: 1700038071230051.png (879 KB, 1024x1024)

879 KB PNG

The usual "anime screencap (frame)" from dalle3 works just fine with flux btw, got this low quality gen from just that prompt

>>101676186
flux seems to blur the pictures from time to time for some reason

Honestly, Dev is impressive, but i'm actually more impressed by how fast Schnell is with the quality of its gens. Even text is usually really good, you can easily gen tons of logos for example.

>>101676203
Nono, this is intended, since the prompt is "anime screencap (frame)" and of course there are some old anime pics in the dataset

>>101676070
maybe a quantised model will fit and not produce garbage, but don't hold your breath

>>101676207
both are distilled no? that means both work at low steps?

>>101676219
there's some instance of people having blurred pictures for no reason though
>>101675193
>>101675235

>>101676203
It's a bug in example workflow. You need to set cfg to 1 for it to work

>>101676253
but cfg = 1 means no negative prompt, I don't want that :(

>>101676277
Nta, have you tried what he suggests anyway to see if the model still gives you what you want without negatives?

>>101676255
>>101676265
Thanks I hate it.

File: 1713722953210749.png (615 KB, 1024x1024)

615 KB PNG

its not a bad model, the finetunes are going to be insane

it doesn't look good for SAI
I doubt they'll be getting another bail out

File: [euler_ancestral_cfg_pp,n(...).png (340 KB, 496x600)

340 KB PNG

Is it broadly correct to say that lowish CFG + high steps count in Euler ancestral normal = a minimalist exaggeration what the model deems the essential features of the prompt? (Details in filename, steps = 80)
Her pants length is especially hard to reproduce (via different weightings of "pants", "shorts" tokens), which makes it a good testbed for behaviour. They either collapse into full length pants or regular short shorts with different relative weights, the window for an intermediate result is very narrow indeed.

By contrast, regular Euler completely ignored some parts of the prompt I deemed essential in the prompt (like her bag becomes a messenger bag instead of a school bag). Euler and its closely related sampling methods are more cavalier in this regard.

People probably figured this out 2 years ago, but I found it intredasting.

File: de_fl_00073_.jpg (894 KB, 1344x768)

894 KB JPG

>>101676411
iirc finetunes for pixart were super slow because hardware reqs were too high to train for the model. I assume flux will experience similar issues. sd1.5/sdxl being more accessible to training was really big for moving those architectures forward

>>101676429
I voted for biden just so I could get the card. does anyone know if kamala votes renew my card or not?

>101676511
>101675700

File: file.jpg (316 KB, 1792x1024)

316 KB JPG

>7465214016
little better
i forgot to eat
gonna go maccies and leave it compile, ill push when im back assuming there's no compilation issues
going to only usable on 32gb gpus though until i implement fp8 or do some sequential workspace or something idk. just for fun anyway, use pytorch, it has more than 1 dev (me)

>>101676485
>>101676485
>>101676485

File: de_fl_00074_.jpg (800 KB, 1344x768)

800 KB JPG

I get my daily rolls for suno
I get my daily rolls for udio
I get my daily rolls for flux
I get my daily rolls for luma
genai really is just gacha gaming. which banner should I be saving for?

File: file.jpg (466 KB, 1024x1792)

466 KB JPG

>still compiling
need to sort out compilation times too
too many files, for some ops it ends up with the same kernel repeated dozens of times

File: 1722568637688_image.jpg (157 KB, 964x1446)

157 KB JPG

Filling

File: 1722569998449.jpg (116 KB, 512x512)

116 KB JPG

Filled!