I am debo - mega cool and zany xdd

File: FDG_News_00005_.jpg (1.04 MB, 1344x768)

1.04 MB JPG

>mfw Resource news

08/02/2024

>Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generation
https://yixiaowang7.github.io/OptTrajDiff_Page

>UniTalker: Scaling up Audio-Driven 3D Facial Animation through A Unified Model
https://github.com/X-niper/UniTalker

>Smoothed Energy Guidance for SDXL
https://github.com/SusungHong/SEG-SDXL

>Mitigating Multilingual Hallucination in Large Vision-Language Models
https://github.com/ssmisya/MHR

>GalleryGPT: Analyzing Paintings with Large Multimodal Models
https://github.com/steven640pixel/GalleryGPT

>The Manga Whisperer: Automatically Generating Transcriptions for Comics
https://github.com/ragavsachdeva/magi

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

>mfw Research news

08/02/2024

>MM-Vet v2: Benchmark to Evaluate Large Multimodal Models for Integrated Capabilities
https://arxiv.org/abs/2408.00765

>Text-Guided Video Masked Autoencoder
https://arxiv.org/abs/2408.00759

>TurboEdit: Text-Based Image Editing Using Few-Step Diffusion Models
https://turboedit-paper.github.io

>SAM 2: Segment Anything in Images and Videos
https://arxiv.org/abs/2408.00714

>MotionFix: Text-Driven 3D Human Motion Editing
https://arxiv.org/abs/2408.00712

>Synthetic dual image generation for reduction of labeling efforts in semantic segmentation of micrographs with a customized metric function
https://arxiv.org/abs/2408.00707

>Scaling Backwards: Minimal Synthetic Pre-training?
https://arxiv.org/abs/2408.00677

>SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement
https://arxiv.org/abs/2408.00653

>Are Bigger Encoders Always Better in VLMs?
https://arxiv.org/abs/2408.00620

>Alleviating Hallucination in Large VLMs with Active Retrieval Augmentation
https://arxiv.org/abs/2408.00555

>Illustrating Classic Brazilian Books using a T2I Diffusion Model
https://arxiv.org/abs/2408.00544

>Reenact Anything: Semantic Video Motion Transfer Using Motion-Textual Inversion
https://arxiv.org/abs/2408.00458

>Towards Reliable Advertising Image Generation Using Human Feedback
https://arxiv.org/abs/2408.00418

>A Simple Background Augmentation Method for Object Detection with Diffusion Model
https://arxiv.org/abs/2408.00350

>ADBM: Adversarial diffusion bridge model for reliable adversarial purification
https://arxiv.org/abs/2408.00315

>Navigating T2I Generative Bias across Indic Languages
https://arxiv.org/abs/2408.00283

>WAS: Dataset and Methods for Artistic Text Segmentation
https://arxiv.org/abs/2408.00106

>Replication in Visual Diffusion Models: A Survey and Outlook
https://arxiv.org/abs/2408.00001

>EmoTalk3D: High-Fidelity Free-View Synthesis of Emotional 3D Talking Head
https://arxiv.org/abs/2408.00297

File: 1664483242244275.jpg (3.09 MB, 4000x2390)

3.09 MB JPG

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.

Is this a good idea, and can anyone tell me very specifically what programs to use and which skills to learn. Are the above steps right?

>>101693292
No, it's not a good idea. Go outside and take your own pictures. If you're too ugly for your own pictures, hit the gym.

File: file.jpg (438 KB, 1792x1024)

438 KB JPG

reposting in case anyone cares

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

also reached 1.1b gens

icl getting zero recognition for anything i do is starting to get to me

tl;dr nothing important

File: Ellie-Ensley @_ellietheem(...).jpg (106 KB, 984x984)

106 KB JPG

>>101693167
>Train LoRA on SDXL
>Net dim:16, net alpha:8
>Training data is a dark skinned woman
>Inference outputs almost always suck ass
>Train again with net dim and alpha set to 32
>Inference outputs still suck ass
>Set them both to 64 and toss it in the oven again
>Outputs look good this time (see pic rel)

I think this may be indicator of slight bias with SDXL. The fact that I had to turn the net dim up implies that it was having trouble learning her face. I've done multiple person LoRAs in the past but both of those were on lighter-skinned people. I think going forward I will have to keep this in mind. Less dark skinned people in the data set seems to be affecting how easy it is to train LoRAs on them too. I don't particularly like that I had to increase the net dim value because this meant the LoRA network file ended up being well over 400 MB. It seems turning the net dim up increases the ease of training but at the cost of having your output files being kind of huge.

>>101693306
I'm already gorgeous and fit as fuck. And I'm not so vapid that I'll do what all the aspiring failed influencers do and fuck up a public setting in order to take two dozen pictures of myself in front of the town square statue.

>>101693306
based

>>101693361
You will literally get more matches with two basic pictures: you with abs lifting up your shirt and you with a dog.

>>101693378
I don't have a dog. That's the whole fucking point of my comments. I'm going to generate pictures of myself with a dog.

> Imagine you had never eaten ice cream in your life-

> But I have eaten ice cream. Why do you keep saying I haven't eaten ice cream?

If someone wants to try schnell/dev but has no VRAM or is just bored: >>101693452

>>101693424
They will ask about the dog. So either find one you know in IRL or don't do it. Girls are people, anon. But let me know how well the socially inept sociopathic pathological lying goes. But sure, why not go all in at that point? Generate some pictures with you with hot girls, $100k cars and $5000/night hotel rooms.

>>101693457
nice try, fed

>>101693464
the NSFW checker is also disabled, yeah

File: file.png (1.2 MB, 864x1280)

1.2 MB PNG

>>101693462
Go all in anon, I believe.

File: de_fl_00009_.jpg (172 KB, 896x1088)

172 KB JPG

>>101693424
>wow anon, your dog is so cute! how old is he?
so, what then, you just keep lying? whats even the point of dating if everything you're trying to sell is a lie?

>>101693457
awesome

File: file.png (1.84 MB, 864x1280)

1.84 MB PNG

>>101693493
Yeah, the car is in the shop

File: file.png (1.35 MB, 864x1280)

1.35 MB PNG

>>101693544
lol my dad is soo cheap

>>101693510
>>101693462
Goddamn you people are a special brand of autistic mixed with retarded. I'm not actually going to photoshop a dog into my pictures.

I'm going to do shit like make a GOOD photo of myself in front of the Eifel Tower, which I have visited. Or at the Sofia Reina. Or just on a fucking bridge over a creek in the middle of nowhere. The point is to make pictures of me that look good. Not alter my appearance or pretend like I've done shit I haven't done.

File: file.png (1.21 MB, 864x1280)

1.21 MB PNG

>>101693593
Just me and my boy

File: Screenshot 2024-08-02 161306.png (2.32 MB, 1386x1746)

2.32 MB PNG

getting nice results with 768x768 atm without slowing my pc to a crawl

File: de_fl_00010_.jpg (320 KB, 896x1088)

320 KB JPG

>>101693655
>only an AI can capture the REAL me
>but you're the retards!
good luck with your catfishing, I guess

>>101693658
>>101693749
score_9, score_8_up, score_7_up, score_6_up, score_5_up, source_cartoon, BREAK, 1girl(Kim Possible),doggy style,leg spreader,bound hands,ball gag,black stockings, lace bra,animal(dog),dog cock,dog penis,canine penis, on stage,large crowd,spectacle,size difference,giant dog,<lora:Age Slider V2_alpha1.0_rank4_noxattn_last:-2.5> Negative prompt: score_4,score_3,score_2 Steps: 40, Sampler: DPM++ 3M SDE, Schedule type: Exponential, CFG scale: 4.5, Seed: 2355056406, Size: 1280x768, Model hash: 821aa5537f, Model: autismmixSDXL_autismmixPony, Denoising strength: 0.7, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer model 2nd: lips_v1.pt, ADetailer confidence 2nd: 0.3, ADetailer dilate erode 2nd: 4, ADetailer mask blur 2nd: 4, ADetailer denoising strength 2nd: 0.4, ADetailer inpaint only masked 2nd: True, ADetailer inpaint padding 2nd: 32, ADetailer model 3rd: hand_yolov8n.pt, ADetailer confidence 3rd: 0.3, ADetailer dilate erode 3rd: 4, ADetailer mask blur 3rd: 4, ADetailer denoising strength 3rd: 0.4, ADetailer inpaint only masked 3rd: True, ADetailer inpaint padding 3rd: 32, ADetailer version: 24.6.0, Hires upscale: 1.25, Hires upscaler: 4x-UltraSharp, Lora hashes: "Age Slider V2_alpha1.0_rank4_noxattn_last: 6da9e7daedf2", Version: v1.10.1

File: 000000_15721_.png (2.51 MB, 1082x1581)

2.51 MB PNG

>>101693216
>>Kolors ipadapter FaceID Plus
>https://github.com/Kwai-Kolors/Kolors/tree/master/ipadapter_FaceID

>Waits for ComfyUi implementation..ppaahhleeezzz

>>101693845
>I'm sorry Policy Officer

File: FLUX__00008_.png (1.3 MB, 1024x1024)

1.3 MB PNG

>>101693853
you will be

free gens for everyone, even with dev! especially if you're a vramlet and want to try flux. >>101693452
flux wins

>>101693925
what's the catch

File: out-0.jpg (277 KB, 1024x1024)

277 KB JPG

>>101693925
Thanks

>>101693937
nta, but he's probably logging, maybe even with IPs, apart from that there's nothing he could do

>>101693937
Your IP, prompts and images are logged

File: w01.jpg (184 KB, 1344x768)

184 KB JPG

>>101693925
thx

>>101693292
Been there, done that (with photoshop, etc...) It doesn't work like you expect unless they're pictures that look natural (show off your body, natural lighting, etc...) and if you're true to yourself then it's useless. You're going to suck at making the images looking candid, etc... so it's not worth it. You're wasting your time with dating apps, better to join a gym and gymmax, then any real photo you take of yourself regardless of how shitty it is is likely to get matches.

File: ComfyUI_10228_.png (1.21 MB, 1440x1120)

1.21 MB PNG

thank god i bought pizza yesterday

>>101693994
Nice, just finished airfrying potatoes and nuggets, w/bullseye sauce and mayo!

File: 00110-1103652947.jpg (989 KB, 2328x1304)

989 KB JPG

ordered lamb saag and regret, very dull taste. should have gotten the classic chicken 65 as usual. and also got (free) eddies for first time in like 8 months

>>101694256
does she give blowjobs?

>>101694256
not your blog, fagoff

>>101694271
id ask >>101694276 's ma

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

1003 KB PNG

>>101694276
Steak tacos with homemade Pico and chimichiri, tater tots with taco seasoning, roasted red pepper salsa

File: file.jpg (339 KB, 1792x1024)

339 KB JPG

added a couple things ready for multi gpu support. idk why im doing this, its not really what i want to do. idk why im doing any of this desu

File: fgdfgdfdgfgd.jpg (419 KB, 1280x768)

419 KB JPG

>>101694349
gosh i hope the taters were crispy
what um how do you guys light your room while genning, certain types of light? i feel i have been using poor quality lighting and its having an effect on my gen quality!

File: xl empty latent.gif (752 KB, 512x512)

752 KB GIF

Why my animatediff render comes so degraded?

>>101694425
state of local

>>101694451
I can't see, move!!

>>101694391
Desk lamp
TV is usually on
Room light off
Hallway light on
I don't have led lights like some of my friends

>>101694468
dalle could never

File: FLUX__00020_.png (1.37 MB, 1024x1024)

1.37 MB PNG

all this potential and I'm still 1girling
crying shame

File: de_fl_00001_.jpg (440 KB, 1344x768)

440 KB JPG

anyone know the token limit on flux

>>101694527
512 I think

File: a2449805d9f56e62560d64eb8(...).png (360 KB, 900x1200)

360 KB PNG

I humbly ask for your honest attempts in replicating this picture.

I want to gauge how far this stuff has gotten

File: loser.jpg (2.21 MB, 3840x2160)

2.21 MB JPG

>did separate Koikatsu renders of Kana+table+chair+shoes and Akane
>did two separate Kana renders - one with feet for size reference and one with transparent feet
>pasted desired feet (using size reference) over render with transparent feet
>img2img+inpaint+canny on the Kana+feet image, repeatedly
>generate just the socked foot separately, copy/paste into Kana image, inpaint over edges so it's seamless
>img2img+inpaint+canny on the Akane render, repeatedly
>copy/paste Akane into the Kana image, delete parts of her behind the table/Kana's foot
>inpaint over edges so it's seamless
>final touchups/edits
Maybe someday we'll have a model with prompt comprehension advanced enough to generate two specific characters in two specific outfits and have one licking the other's foot.
We do not have that model, so workarounds are required.

File: file.png (3.45 MB, 2527x1239)

3.45 MB PNG

>>101694380
Comfy already supports multi-gpu?

File: 1718676901864225.png (1.06 MB, 1024x1024)

1.06 MB PNG

flux is neat so far, text works, logos work

>A cute blonde anime girl wearing a white Edmonton Oilers hoodie standing on the street holding a rectangular white sign with the word ANIME written on it in black text

>>101694425
Try webm not gif

File: 00128-2477747725.jpg (1.02 MB, 1176x4000)

1.02 MB JPG

kinda insane how 1.5 (actually) is (still) the best,

>>101694656
also note if you dont have 12+ gb of vram or even a on a 4080 you need to use the fp8 clip model

>>101694687
Cool

File: 25026_autogen.jpg (416 KB, 1752x1336)

416 KB JPG

>>101694745
Nice tunnel

And the award for worst LORA ever goes to: https://civitai.com/models/595059/diversify-sd3

File: 1711050222717624.png (993 KB, 1024x1024)

993 KB PNG

every day tech gets better and better.

>>101694786
by that logic, use it in negative
it is now the best lora ever

>>101694726
But it's not for anime girls and that's all that matters?

>nobody has made an underalterbach lora yet

>>101694806
is this e.s.l.

>>101694797
>it is now the best lora ever
It's still an SD3 LORA, anon.

>>101694628
cool but im not working on comfy

>>101694823
No?

>>101694791
I really hate that Twitter is the place with the most art and the place where artists get the most views because it is absolutely NOT designed well in terms of finding art.

File: de_fl_00005_.jpg (413 KB, 1344x768)

413 KB JPG

>>101694855
twitter is so shit. I made an account just to see happenings in the AI space but have to aggressively prune my algo every other week cuz it keeps trying to feed me weird nazi shit or clickbait trash

File: 1720581557675726.png (970 KB, 1024x1024)

970 KB PNG

>>101694791

>>101694908
>I made an account just to see happenings in the AI space
kek just follow /sdg/ /lmg/ and /aicg/
They have all the news.

File: 1710360267406807.png (1.17 MB, 1024x1024)

1.17 MB PNG

>>101694926

koff is based

>>101694908
Need this with teen prostitute X-23

16 bit Street Walker beat em up

>>101694947
why do the rocks look so repeating

File: de_fl_00007_.jpg (294 KB, 1344x768)

294 KB JPG

>>101694937
I'm the news guy lul. I have to be on twitter to be the first to see stuff

>>101694996
here's a title screen

>>101695045
Ah. Thank you for suffering from Twitter brainrot for the benefit of the rest of us.
I try to expose myself to a little of social media as possible.

File: file.png (1.82 MB, 1024x1024)

1.82 MB PNG

>>101694380
want do you want to do then>

>>101695045
Lol, awesome

File: 00146-1304381683.jpg (526 KB, 1280x2064)

526 KB JPG

i walk these streets, at night
hey buddy, got a light?

>>101695072
data but nobody is interested in that either

has there been any news on a1111 or forge getting flux up and running? I have it on comfy but I just don't like comfy and want things like in-painting done on the a1111 style

File: 1716067060096387.png (1.17 MB, 1024x1024)

1.17 MB PNG

the text related prompts are good in flux, even for background art:

File: ComfyUI_01573_.png (1.23 MB, 896x1152)

1.23 MB PNG

>>101695145
just people complaining it's not in

File: file.png (2.12 MB, 1280x768)

2.12 MB PNG

>>101695133
I'm interested

>>101695162
you're probably only interested in what i can do with the data rather than using it yourself

File: 1708371733301911.png (1.26 MB, 1024x1024)

1.26 MB PNG

>>101695151

mfw you see a great artist and think to maybe make a lora but their bio says that is prohibited.. frustrating
captcha attempt number 3..

FUCKEM'

File: sataniaanyaface.png (2.47 MB, 2048x2048)

2.47 MB PNG

>>101695233
Do it anyways.

File: de_fl_00008_.jpg (449 KB, 1344x768)

449 KB JPG

>>101695233
>but their bio says that is prohibited
damn. air tight defense.

File: ComfyUI_00116_.png (1.13 MB, 1024x1024)

1.13 MB PNG

a photo of a fox with a sign that says "hello 4chan" in it's mouth. The fox is in an open field with grass and flowers. The sign is made of wood and the fox is biting in it's mouth. This is a normal fox and not anthropomorphic in any way.


Any suggestions for making the fox look more real and not cartoony?

File: out-0 (15).jpg (370 KB, 1024x1024)

370 KB JPG

Holy shit Flux is cool

>>101695388
try a sort of Oscar the grouch creature in a toilet bowl holding a sigh saying 'not your blog'

File: 1_5 empty latent.gif (1.29 MB, 512x512)

1.29 MB GIF

>AI can't still make animated sprites
This tech is a scam.

File: ComfyUI_01579_.png (778 KB, 1152x832)

778 KB PNG

>>101695431
just fake game screenshots to scam people in preorders

File: file.png (1.9 MB, 1280x768)

1.9 MB PNG

Any good ryona prompts?

>>101695423
fuck off redditor

>>101695388
>and not anthropomorphic in any way
don't use counter-factuals. you're just adding 'anthropomorphic' into the encoding

>>101695371
Nice

File: de_fl_00009_.jpg (465 KB, 1344x768)

465 KB JPG

>>101695455
how are you prompting for pixel art? I can't get it to listen to me

File: ComfyUI_01575_.png (1015 KB, 1152x896)

1015 KB PNG

>>101695516
you have to describe it as a game not a screenshot
>a beat em up arcade game catgirl with long grey hair, grey eyes addidas tracksuit, yelling, pixel art side scroller

File: de_fl_00011_.jpg (315 KB, 1344x768)

315 KB JPG

>>101695536
>1990s side-scroller game in the style of Castlevania with a junky urban environment, the main character is teen prostitute X-23, retro pixelated RPG game side-scroller, pixel art
hrmm

>>101695573
Best one so far

File: ComfyUI_01586_.png (706 KB, 1152x896)

706 KB PNG

>>101695573
what are your settings?

File: de_fl_00012_.jpg (367 KB, 1344x768)

367 KB JPG

>>101695599
boss fight! even put her name on the health bar

>>101695632
I'm just using anon's hosted webapp rn. none of the settings are surfaced

>>101695671
Red hulk let himself go

File: ComfyUI_01491_.png (989 KB, 1024x1024)

989 KB PNG

>>101695671
you can run it on 8GB apparently but idk how patient you are

File: de_fl_00013_.jpg (350 KB, 1344x768)

350 KB JPG

>>101695700
yeah I've been meaning to get it set up on my machine but I've been colossally lazy today... I guess no time like the present

File: out-0 (17).jpg (318 KB, 1024x1024)

318 KB JPG

>>101695573
>pixel art side scroller, side-scroller game in the style of Castlevania with a junky urban environment, the main character is teen prostitute X-23, retro pixelated RPG game side-scroller, pixel art


>CFG 1, Dev, AR 1:1

flux can do a lot of stuff like dall-e including text and brand logos, I think sdxl + loras + controlnets are best for characters, but this can do a lot of neat stuff so far and i'm just starting out with it.

>>101695431
run it through the pixelize extras option in auto1111

https://github.com/AUTOMATIC1111/stable-diffusion-webui-pixelization

batch process the video and poof

File: de_fl_00014_.jpg (357 KB, 1344x768)

357 KB JPG

wtf, surprise debo. I didn't even prompt for him, just "boss fight with vampire"

File: file.jpg (393 KB, 1792x1024)

393 KB JPG

every transformer_block and single_transformer_block is the same, i could just split the model up and it would run in however much vram is required for 1 block

>>101695748
>>101695766
These are great, thanks for the gens

File: 1722650292.jpg (210 KB, 1024x1024)

210 KB JPG

>>101693925
thanks

>>101695766
>>101695748
I'm wondering if perhaps the landscape aspect ratio is making it want to draw an adventure game vista rather than low res dot pixel art

>>101695830
Maybe it has to be 4:3 like old tvs?

File: kanaanyaface.png (2.87 MB, 2048x2048)

2.87 MB PNG

https://civitai.com/models/378589/anyas-heh-face-meme-or-concept-lora-xl this is my all-time favorite LORA.

>>101694786
cool. there are SD3 finetunes now
https://civitai.com/models/564563?modelVersionId=667396

>>101695858
I downloaded Aqua cry face and yell face from konosuba but haven't used them yet

File: de_fl_00015_.jpg (392 KB, 1024x1024)

392 KB JPG

>>101695819
np, its a fun prompt

>>101695830
>>101695841
tried a few different ratios but it doesn't seem to change up the style

Anyone tried this yet? https://huggingface.co/Kijai/flux-fp8/tree/main

>>101696016
these are what you use if you dont have a 4090 or 24gb VRAM. I can't even use the 23gb ones on a 4080.

>>101696014
Lol she has her own office

Thanks for the gens, I gotta go for a bit

File: 1695422319561032.png (1.35 MB, 1024x1024)

1.35 MB PNG

fp8 model:

>Akihabara Japan, lots of electronics shops with neon signs with "SDG" on them, street signs with "SDG" in black text

>>101696016
>>101696035
its just a smaller file. it doesn't affect memory usage or gen speeds at all

>>101696016
Yes. This is in e4m3fn format btw. Make sure to run it with that setting in Comfy.
It outputs the exactly the same the thing as the original fp16 weights set to fp8_e4m3fn.

File: 1691349181983744.png (1.88 MB, 1024x1024)

1.88 MB PNG

>>101696069

>>101696121
is the quality the same? i'd assume better results with the large file

>>101696175
there's no difference between running the full model in fp8 or running the fp8 model in fp8. its just a smaller file because the weight precision is pre-trimmed, rather than trimming during run time. there's no computational difference

>>101696175
>>101696233
this still means that the fp8 model is worse than the original fp16 one

>>101696237
I don't think anyone is arguing otherwise

File: FLUX__00050_.png (1.36 MB, 1024x1024)

1.36 MB PNG

I asked for quote "large breasts", and this is what I got

File: file.jpg (309 KB, 1024x1792)

309 KB JPG

<aitemplate.compiler.transform.memory_planning> Workspace shared_size=56623104 unique_size=0
<aitemplate.compiler.transform.memory_planning> max_blob=314658880 constant_offset=0
each FluxTransformerBlock would use about 350mb working memory and about fp16 650mb weights, 1gb total
<aitemplate.compiler.transform.memory_planning> Workspace shared_size=56623104 unique_size=0
<aitemplate.compiler.transform.memory_planning> max_blob=452990976 constant_offset=0
then each FluxSingleTransformerBlock would use about 475mb working memory and about fp16 270mb weights
should take about the same time as running the full model in one go would take, wouldnt necessarily need to load the entire model weights to ram either, could be read from disk for each layer

so the whole model can run in ~1gb
i thought of doing this for other models before but the blocks and input shapes for each of sd are different so it was just awkward

File: 00188-1920735951.jpg (866 KB, 2024x1504)

866 KB JPG

this aint your frog, pal

>>101696335
why don't you just make your own dalle3?

File: file.jpg (282 KB, 1792x1024)

282 KB JPG

>>101696359
i dont have to because it already exists

>>101696359
money

>>101696319
it's clearly limited by the dataset in many regards
finetuning is necessary

>>101696319
try "great honkin huge silicon melons"

is sd3 fixed yet?

File: nagatorovomit.jpg (1.16 MB, 1859x2048)

1.16 MB JPG

>>101696319
>women who are fat by virtue of having large breasts

File: I swear this is not an Ad(...).png (2.57 MB, 1624x782)

2.57 MB PNG

>>101693167
Did any of you know you can use Huggingface's "Serverless" Inference API to test LoRAs trained on SDXL 1.0? You can gen images on site, like this page I set up:

huggingface.co/AiAF/Ellie-Ensley__ellietheempress_LoRA_SDXL-1.0

File: delux_mtg_00001_.png (1.95 MB, 896x1152)

1.95 MB PNG

5s/it
~2min per gen
not the worst

i think trani is shit (as a human)

>>101696537
>0 downloads
far too many

File: 1707872250804196.jpg (119 KB, 984x984)

119 KB JPG

>>101695229

>>101693167
australia clock

>>101696519
Nice, was that fp8 on both the t5 and the checkpoint?

File: vomit.png (995 KB, 1825x417)

995 KB PNG

>>101696615

File: delux_mtg_00002_.png (1.7 MB, 896x1152)

1.7 MB PNG

>>101696623
yeah

>>101696519
steps/gpu?

File: delux_mtg_00004_.png (1.67 MB, 896x1152)

1.67 MB PNG

>>101696652
20 steps
4070

File: 00204-1756418954.jpg (1.27 MB, 2528x1880)

1.27 MB JPG

https://youtu.be/aI5ZzQbC8i0?si=nzizLn0g_Gtd3rjj
music is my unwitting internal state

File: delux_mtg_00006_.png (1.79 MB, 896x1152)

1.79 MB PNG

>>101696803
>internal state is dogs barking, musically
I guess that makes sense

File: 000000_15734_.png (1.67 MB, 1024x1365)

1.67 MB PNG

> Flux, t5xxl_fp16, 20 steps, 3060(12G), 64ddr4....9mins gen..

File: delux_mtg_00005_.png (1.78 MB, 896x1152)

1.78 MB PNG

>>101696920
>9mins
oof

>>101696949
>oof
right? //needs 3090Ti...

File: delux_mtg_00007_.png (1.64 MB, 896x1152)

1.64 MB PNG

>>101696955
can just use anon's webapp if you're not afraid of the FBI
https://lodging-traditional-working-form.trycloudflare.com/

File: 1717751179776548.png (947 KB, 1024x1024)

947 KB PNG

switched to the 23gb checkpoint (after a long download) from a 11gb pruned one, outputs are much nicer it seems

clip is fp8 (I only have 32gb physical RAM)

File: 1715964370023468.png (1.07 MB, 1024x1024)

1.07 MB PNG

>>101697094
before, same prompt/steps:

File: mikuquestion2.jpg (989 KB, 1710x1779)

989 KB JPG

>>101697094
Model that consistently outputs symmetric eyelashes when?

>>101697112
not sure

anatomy also seems to work well, text and logos work really well which is neat

I assume at least someone here has Flux running on windows. 4090, 64gb ram, grabbed ComfyUI portable, ran it's update scripts, loaded workflow from the image, all that shit, and it won't work.

Best I can tell, shit crashes the exact instant that the request is put into the prompt queue with a memory access violation by fucked if I know.

So if you have it working on windows, did you use the portable install, or just git cloned it and manual install?

>>101697131
How are hands and feet?

File: 1705559364038805.png (942 KB, 1024x1024)

942 KB PNG

>>101697159
I just started but it passes the test initially

File: file.jpg (280 KB, 1024x1792)

280 KB JPG

ait, 3060 12gb, 1024x1024
70ms for FluxTransformerBlock, * 19
same for FluxSingleTransformerBlock on 3060 12gb, * 38
rest of the model is basically just some linears, layernorms and concatenates, ill check but should be negligible
so maybe ~4s/it, assuming changing weights doesnt add too much time and ~1gb vram

>>101697039
TY but just a local dweller.

File: 1692137780000737.png (1.04 MB, 1024x1024)

1.04 MB PNG

now THIS is podracing.

File: alyagarbage2.jpg (639 KB, 1446x1080)

639 KB JPG

>>101697150
>windows

>>101697192
Sorry, some of us have actual work to do.

>>101697200
I've been working at home on Linux for years.

File: file.png (2 KB, 346x54)

2 KB PNG

gets around the 2gb bound constant limit too
entire model could be released as modules with weights, bound constant modules are faster to load than applying constants after

File: tmpvi90rd4q.png (658 KB, 768x768)

658 KB PNG

>>101697150
>shit crashes
As in, the program crashes or a BSOD or what?

File: 1711610721217613.png (1014 KB, 1024x1024)

1014 KB PNG

>>101697183

>>101697150
portable, fp8 clip version (dont have 24gb vram), 32gb ram, 4080, works fine

used the zip didnt git clone anything.

File: 1712623764490735.png (1.08 MB, 1024x1024)

1.08 MB PNG

ok, now we're talking.

>>101697239
"got prompt" and then nothing. Seems like it doesn't even try to load the model, no vram movement at all, no ram usage, nothing. Just puts the prompt into the prompt queue and program exits.
(Line 507 of server.py)
PyCharm's debug window tries to refresh the variable view after that line returns and just gives up.

File: delux_mtg_00009_.png (1.86 MB, 896x1152)

1.86 MB PNG

this one is kind of koff-coded

>>101697167
*nods agreeingly, clearly not understanding anything you just said*
sounds promising

>>101697238
right, i forgot, no contributions allowed, nobody cares about running the model everyone is talking about faster and with way less resources

File: 1717236065384930.png (994 KB, 1024x1024)

994 KB PNG

>>101697280
also the variety is good, you could even use this model for img2img or inpainting in theory to do custom text or backgrounds, with 1.5 or SDXL generations.

File: tmpxq8rzcqv.png (945 KB, 768x1024)

945 KB PNG

>>101697250
>>101697280
>>101697311
Why is she on a chair outside?

File: delux_mtg_00010_.png (1.43 MB, 896x1152)

1.43 MB PNG

>>101697296
interesting perspective, but have you considered fucking off instead?

File: 00225-217086121.jpg (907 KB, 2328x1656)

907 KB JPG

this aint your blog, maaaan!

>>101697167
i bet you're trans

>debo

File: file.jpg (253 KB, 1024x1792)

253 KB JPG

>>101697285
yeah it is promising but i'm fed up of getting shit for everything i do. i'm going to bed. wtf is the point lmao

>>101697341
well I said sitting outside on a chair but I meant sitting on a chair outside the entrance.

File: ComfyUI_170114_.png (1.06 MB, 1024x1024)

1.06 MB PNG

>>101697282
If you are using the latest standalone it means there's something wrong with your system. Do you have an 13900k or 14900k by any chance?

File: delux_mtg_00014_.png (1.52 MB, 896x1152)

1.52 MB PNG

>>101697376
you can't let the nogens win :(
but I get it. gn, be well

>>101697238
>>101697296
>>101697390
lmao

File: ComfyUI_30504_.png (1.06 MB, 1024x1024)

1.06 MB PNG

>wanted to check weapon generation
>added "holding desert eagle"

>>101697450
need a barn owl

>>101697450
Dps might be lower but at least you won't need to reload

File: file.jpg (334 KB, 1024x1792)

334 KB JPG

based janny
>>101697410
its fine ive just had a bad day. still i dont need to put up with this shit. im ok with local flux being slow af, i use dalle anyway

>>101697513
kys tranny faggot, you will never contribute anything of value in your life

File: 1716512422910686.png (1.03 MB, 1024x1024)

1.03 MB PNG

goku punching peter griffin, it gave them both the family guy style

>>101697513
>i use dalle anyway
kek

>>101697519
why be such a miserable shitty person

File: de_fl_00101_.jpg (1.5 MB, 1344x960)

1.5 MB JPG

dang, the webapp is offline
he giveth, and he taketh

>>101697542
oh, you reminded me to try my challenge prompt
>goku powerbombs hatsune miku through the tournament floor at the tenkaichi budokai
no model has gotten it yet

File: 1696315539431682.png (954 KB, 1024x1024)

954 KB PNG

>>101697628
this is what I got for miku with an ak47 and camo dress:

>>101697538
>>101697538
>>101697538

File: tmphxlcax_g.png (1.26 MB, 768x768)

1.26 MB PNG

>>101697613
Because anonymity is a hell of a drug