Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101927580>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>GPU performancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>Maintain thread qualityhttps://rentry.org/debo>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
Does flux know what semen is? I tried a lot of tags and only thing that kind of works is "white liquid".
>>101930764Pretty cool.
>>101930742>my slop is in the rentryblessed
Im starting to notice my refiner might be fucking up my gens. For example i'm using autismmix here, so.. Should i still be using it as the refiner or is there like a designated model for that purpose?https://files.catbox.moe/lfp5zk.png>screenshot is of the 2nd in the batch btw
All "realism" posters must be stopped. Plastic skin is not realism, do us all a favor and stop generating synthetic 1girl slop.
>>101930814I like it though.
>>101930848pedo spotted
>>101930814>All "realism" posters must be stopped. Plastic skin is not realism, do us all a favor and stop generating synthetic 1girl slop.thank you for the prompt 8/10
Last night I slept only 5 hours because of this shit.
>>101930774mate it hasn't even seen a female nipple, why would it know what semen looks like
>>101930900just poisoned, ive gotten titties before. Set steps to 5 and youll see a lot of shit you prompt
>>101930915catbox one passable nipple from base Flux, I dare you
>>101930830>>101930938these are nice
>>101930929I've seen it generate plenty of male and little girl nipples no problem. It just can't place them on a boob.
>>101930945fuck off teebs
>>101930895
>>101930915That has been my take on the model. New and cool, completely worthless for my use case.Prompt: voluptuous blonde girl looking down on me, absurdly huge breasts barely contained by shirt, tall, cute, from low view, garage exteriorThis is what flux makes for me
>>101930814>>101930891vouch prompt is solid 8/10
>>101930943why, thank you
>>101930958
>>101930929Behold, blurry 1girls.https://files.catbox.moe/mlcq58.png
>>101930981yeah you nailed the realism look here, crazy good prompt
>>101930965This is what a year old model makes for me.same prooooompt>voluptuous blonde girl looking down on me, absurdly huge breasts barely contained by shirt, tall, cute, from low view, garage exterior,I'll wait until there is good lora support and someone figures out how to undo the poisoning on boobs.
>>101931002I completely slept on SDXL. I downloaded Pony Realism, but it seems complicated to set up. Is it worth it on 8 GB?
>>101930987
>>101931017it's <7GB and has everything included, there is nothing to set up
>>101930965One anon had a lot of success with ridiculous proportions in this thread.https://desuarchive.org/g/thread/101879426/
>>101931031
>>101931048Here, didn't mean to post archive link. >>101881206
>>101930895same dude, I can't sleep well anymore I just want to wake up to to do genMaxxing again ;-;
>>101931045I meant the whole refiner thing. The ComfyUI "simple example" is massive. And Forge crashes on everything.
>face looks perfect in the blurry preview thing on forge>dozen steps later shits out this
>>101931076no one uses the refiner, it was forgotten soon after release, you can delete that whole section from the workflow
>>101931070>I just want to genMaxxingWhat for? It's cool to generate porn and all, but I don't get this obsession with generation. How many hyperspecific pictures do you need?
>>101931017I don't know anon. I use epicrealism_naturalSin because it makes naughty pictures with really big boobs and can do perspective for my giantessfag needs.Otherwise I'm just a coomer with a nice graphics card.I can't tell you anything>>101931048I'll check it out, thanks anon
>>101931124>How many hyperspecific pictures do you need?Several.
>>101930929https://files.catbox.moe/xbgyo2.png
>>101931138holy fuvk
>>101931124I don't know man it's just fun to have something in mind and flux nailing that shit almost everytime, I feel like that NPC from persona 4 who can't stop polishing his basketball balls kekhttps://www.youtube.com/watch?v=xFhS0sSMys8
>>101931146chewy
>>101931138Those are rooki-oh, carry on then
>>101931096OK, I'll try to generate an image now...oh. 42s/it. Yeah. This is going to be fun.
>>101931161what's your hardware?
>>101931167sigh... Here we go again.RX 5700 XT (8 GB VRAM) + Ryzen 5 3600X 6-Core (56 GB RAM) on Linux.>inb4 my condolencesYeah, yeah.
>>101931167PC
>>101931187you should really buy an xbox if you're going to image gen.
So, what is the difference between this thread and the /sdg/ one?
>>101930814
>>101931186my condol-56GB? what.
>>101931203no avatarfags vs avatarfags
>>101931138I have 75k files in mine but only 62GB. Wtf is in that folder? Are you including the programs and models themselves?
Why does SDXL do this?>>101931217I used to rely on CPU + RAM for textgen.
>>101931203This is for local diffusion models generally, stable diffusion general pertains only to stable diffusion. When asked to change the name to reflect the growing market of alternatives, their was a split due to people not accepting that generalization would be better for the long-term development of image diffusion models. The /sdg/ threads keep being created despite the majority of the general members now being here.Take it up with them.
>>101931237SDXL was trained for 1024x1024
>>101931193this one?
>>101931237>I used to rely on CPU + RAM for textgenGod, how slow was it?
Architectbros, flux status?
>>101931232I have a lot of (failed) local video-gens in here.
>>101931138>10k foldersOk yeah that includes your installs lmao.
At the end of the day thread splitting and general autistic flinging that fills the catalogue only comes from the fact that chinkmoot and his gang of deep state glow niggers refuse to give us an /ai/ board.
>>101931256
>>101931078God I wish I could actually prompt for jacked up hooked noses like that.
>>101931247Well, I'm fucked then. Alright. I'm going to bed.>>101931255Not that much when it comes to inference (with 7B models). Context processing was a pain, though. I since managed to run mini-magnum (12B) with HIPBLAS and never looked back.
>>101931245>stable diffusion general pertains only to stable diffusionYou'd think so but they are all using flux too
>>101931265>2 threads openvs>a gazillion threads to hunt through for info and genseven 2 is too much sometimesand that's not even counting the dozen or so from other boards
>>101931282hey i know that img2img!
so i removed refiner (thanks >>101931096 )and my shit is still getting fucked up when it reaches the upscale stagepls someone help these gens look like they could be really goodhttps://files.catbox.moe/accsz5.png
>Forge added official gguf support>Comfy is sleeping on itOh shit I might go back after all.
>>101931286Yeah, you can see the hypocrisy. They should start posting here.
>>101931292lower your denoise < 0.05
I'm starting to think that in 'anime mode' text tends to get fucked up. No problems with realism. that's it.
>>101931295Forge is incapable of managing memory properly.
>>101931295Why should he when there's a custom node from the guy who made those gguf models?
>>101931291Heh, it's not img2img though just a description of it in txt2img.
>>101931295Just use the custom node made by the developer of the GGUF models, anon. Surely they'd have the best implementation.https://github.com/city96/ComfyUI-GGUF
>>101931309No thanks because of >>101931227
>>101931328works on my machine
>>101931332Oh shit, very nice. Fooled me.
>>101931335I've tried this and it gave me an OOM. Even an fp16 runs fine so I'm not sure why a q8/q5_1 would fail.
>>101931337>no avatarfags>guy who keeps posting his reddit account and black miku on a skateboard>>101930762 whoever the fuck this is>pedononce is basically an avatarfag
>>101931329>>101931335Is he working on a lora support? Or is this a job for comfy (who's just gonna steal code from forge again)
>>101931322giving this a try, but suddenly my gen times exploded to 16.5s/t WTFi havent changed anything in my setup except the denoising strength like you said
>>101931355I had the same problem. Redownloaded the models and now it's fine.
>>101931372that's weird... but I guess I will try it and see
>>101931363Yeah but he's a redditor, you can't expect him to be unretarded. At least he is doing useful testing.Meanwhile the other avatarfags are more retarded.>Oh boy, an anonymous message board where nobody will know who I am>Whelp, time to tell everyone who I am.
>>101931367Why don't you read the github page?>LoRA / Controlnet / etc are currently not supported due to the weights being quantized.
>>101931392And yet it just werks in forge
>>101931371well you're barely changing the image so it doesnt work as much
Gonna freak if Forge abandons their UI and tells to go back to A1111 again
The taef1 support has been implemented on ComfyUihttps://github.com/comfyanonymous/ComfyUI/commit/1770fc77ed91348f4060e3c0b040c1519d6f91d0#commitcomment-145464339that means no more blurry shit on the preview, if you chose "Preview method: TAESD" on the ComfyUi Manager Preview, I'd suggest you to download the files on that github comment as they don't appear automatically during the update somehow
>>101931405no im realizing i just sleepily fucked up my node arrangement here, i did something horribly wrong while arranging and its loading the upscaler before the k sampler..man this shit is excruciating to get used to.>>101931422abandon forge and Anal1111
>>101931402fuck i thought you were trollingcomfy why haven't you copied the code yet
>>101931363Both of the latter avatarfags are from /sdg/
>>101931428>man this shit is excruciating to get used to.welcome to comfyui
>>101931432>what a crazy night of meth
tri-devil is all you need
>>101931459Maybe if at least one of them had some tits I would be tempted to evil
>>101931488>titsNo thanks I'd rather we keep this thread SAFE and FUN
>>101931445Meanwhile be meFind 1 workflow I like and never change it.
>>101931510
>>101931445i genuinely dont know why it was skipping the k sampler, i disconnected the loras and it worked again. fuck. well, this is better than struggling with forge given if forge breaks there's literally no possible way to fix it except a reinstall. here it's just "something is wrong with the node or your setup" and its correctable.
Q5 isn't too bad.
>>101931538>Entirely generic 1girl on a blank backdropYou could make that with SD1.5
flux is such an intimidating tool.Did they really need to do all this gayshit with the bubbles?
>>101931533or comfy continues pushing to main kek
>>101931562I was comparing it to nf4
>>101931538>>101931562>Entirely generic 1girl on a blank backdrop no hands visible no interesting pose or props or anything else going onposts like this should be met with death by stoning>>101931573god dammit id be laughing if that didn't hurt so much, this is why you don't UPDOOOOOT people.
>>101931569I guess that's the price to pay to get dalle at home
>>101931569>a basic node workflow filters nocoders this hardliterally just read it from the load checkpoint to the end, retard
>>101931348Thanks, nice Jinx.
Where the FUCK is Nick Fe?? wth
>>101931590i dont mind the complexity for the future transexuals of /g/, but would it really hurt to just have a flat-ui for the basic bitchers. All we need is the model and a few simple sliders.
>>101931636https://github.com/lllyasviel/stable-diffusion-webui-forge
>SPAMS the Flux .1 D feed with shitty bloated (300mb wtf?) celebrity loras Fuck this guy
>>101931590there's absolutely nothing wrong with having bare basics ready to go, a few sliders and whatnot, not everyone's ready to start their journey on hrt and suicide contemplation bub.
>>101931646looks amazing thanks m8, tired of tranny UIs
>>101931649block him?
>>101931656don't worry, even brainlets get a chance to generate images
>>101931649you should see the size of his XL loras
>>101931562But can SD1.5 do this?
>>101931682no gen?
>>101931683Are they extra large?
>>101931699for you
>>101931683>971MBHoly fucking kek, what is this dude doing
is there a comfyUi node that would tie two parameters values together, like if I put cfg at 6, then the value of the other parameter will change accordingly?
>>101931699the smallest is 300MB, he has done loras over 1GBover 1GB just for a face, his loras suck at full body shots and fall apart. only good for inpainting
>>101931691>no gen calling others a no genhere have some slop from yours truly
>>101931649>>101931683>>101931709i was noticing this earlier while flipping through celeb loras for XL/Pony, it seems specifically the celeb genners have their mental illness dialed to such an extreme that they don't notice glaring problems in EVERY one of their Loras, a particular fella has this very particular issue were every model has their eyes spaced a specific way, regardless of the facial structure, i'm 100% positive if you just jumped on civitai right now and looked around and celeb loras you'd see who i'm talking about, people complained in his comments toocivitai is hilarious like that sometimes.
>>101931720When XL first came out my LoRAs were that huge too. It's embarassing but people still give me buzz for them so I don't delete them.
Any cool t5 quantizations so far?
>>101931742>>101931691
>>101931743there was this guy making tons of Pony loras of obscure characters but all the sample images he used were of them extremely obese (the loras were obese in size)the characters in the loras were normal but few people looked at them because you'd naturally assume they were for BBW enthusiasts so they had very few downloads
>>101931651That's what forge is for. I don't understand people who come to Comfy and start whining that it isn't a simple to use program. The reason why it can do so much shit that other frontends can't do is because it is so open-ended and completely directed by the user.Like genuinely. You want the bare basics, a few sliders and whatnot? Forge.
>>101931851MR ADVERTISER GET DOWN-
>>101931851>>101931835MODS
>>101931432> merge_lora_to_model_weightWhat if you didn't merge the lora into the model weights.
>>101931851MY VIRGIN EYES
>Male Ass - Fluxsum total of degens that use Comfy
get in bruh
>>101931851HELPPP AHHHHH WHAT ARE THOSE THINGS!!!!
>>101931851debo?
Can you retards help me figure out the missing stuff instead of spamming shit?
is there any risk of optimized models creating schisms in finetunes/lora training? i know basically nothing about model structure
>>101931826>Stop arguing about software on /g/!kek
>>101931851because you have the Flux F1 filter ticked, which means everything gets filtered out you based retardgood luck with the vacation
>>101931896You unticked all the g and pg shit didn't you
>>101931902add a realism lora to your flow, bro, shit's plastic
>>101931851O Lord REEKRIS!Help me lawdMODS help or jannies it doesn't matter!!!HELL NO!
>>101931908Yes
>>101931908yeah because I'm not a baby, I don't want to see loras for babies
>>101931742you (((genning))) Gen from Fortnite doesn't prove anything mister
>>101931912the high cfg is part of the aesthetic
>>101931946gnarly
>>101931908Thank you anon, that fixed it.I don't know if I'm retarded or their system to choose what to see is what is absolutely retarded.
>>101931902I will stick to the purest form of love
>>101931946how did you manage to get something like that? kek
>>101931977epic
>>101931974>the purest form of loveanon, I'm a hardcore yuri fan, and a thressome with a dude isn't yuri or isn't "the purest form of love", that's bullshit
>>101931938kek, no gen cope
>>101932014excuse me, this is the language used in /sdg/ don't bring that shit here thanks
>>101932024>>101931691>>101931691>>101931691not the one you need to be talking to fren
that saved my poor eyes thanks
lmao bro posted the equivalent of 20 posts worth of porn in just one post and the jannies still havent snatched it outthat's impressive
>>101932067absolutely killer, catbox possible?
>>101932086workflow might be a bit shit cause i was still experimenting with stuffhttps://files.catbox.moe/eb72za.pngjust grabbing random sentences from places can make some absolutely amazing prompts. here's another gen from that prompt.
>get a comfy catbox from someone>MISSING NODE MISSING NODE MISSING NODE MISSING NODE MISSING NODE MISSING NODE MISSING NODE MISSING NODE MISSING NODE MISSING NODE MISSING NODE Every time.
>>101931837Cool
>>101932146kek, i try to keep mine minimalist but it's cool to explore new nodes and implementations people use. also a lot of them are pretty useful for qol
>>101931851
>>101932083It's too late, I already jerked to it.
>>101932141legend, thank youthats my favorite thing to do as well, just mash a lot of oddball shit together and see what comes out
>>101932141prompt sounds like a quote out of warhammer 4k
>>101932146Even better if it's a custom node and you load the examples and none of them work because they use 100 unspecified other custom nodes
What's the status of StableDiffusion.cpp?I tried running it yesterday using FP32 and then compared it against Q8_0 and inference wasn't any faster.Is it doing the FP32 -> Q8 -> FP32 conversions during the MatMul?I'm a CPU-fag, so not being able to do fast quantized inference is causing me some coping, seething and a little bit of dilating.
>>101932053MORE TITS
reposting because i want to ask in both threads: whats the difference between /sdg/ and /ldg/? dont they cover the same thing?
>>101932306We already answered you
>>101932316but then I'd catch that nasty skin condition
>>101932306/sdg/ had a schism due to alternative models pixart/hyuandit. /ldg/ was created so pixart chang's could have a safespace as they huffed copium that it would usurp SD. /ldg/ was then co-opted by resident schizos as another means of dramaposting about /sdg/ avatarfags. /ldg/ was dying for the most part until Flux released, with most people migrating to /ldg/ to post Flux stuff due to it not being an SD model.
>>101932332i like pointillism
@101932371d*bo moment thx for the bump
I bet DALL-E 3 is like 8B (+5B from T5)
>>101931245I see. But it's for image gen mostly, right?
>>101932394>Figure 1. schizo drama-fagging about /sdg/ avatarfags
>>101932405Only if it is diffusion, buster! we don't want no autoregressive image gen images here!
I won
anyone missing an arm on one of their gens? I think I have it
>>101932466well at least she's not hoarding chromosomes
>>101932434uhhhh huhhhuuhhehhheh
>>101932476good one! high five!
>>101932482>>101932476>>101932466samefag
>>101932491(You)
>>101932482haha lol>>101932491retard
>>101932434
oh no, 2 browsers!
>>101932498fake
What's the best way to get into Flux with 12gb vram?
>>101932369>>101932264these look legit, do anything specific to de-plastic?
>>101932521waiting 6 months for new tech breakthroughs that make it blazing fast on 12gb of vram
>>101932316Have they invented an outside yet where I can scream incoherent ramblings about dynamic poses, intricate clothes and messy hair styles at 1girls and not get pepper sprayed?
>>101932507>>101932513I actually have 7 browsers. I just rotate gens and typing styles to look like different people.This guy? >>101932388Me>>101932255I'm also this guy>>101931880This is me too>>101931837(You) got it>>101931826Telling myself to go touch grass, haha>>101931686I have a yuri fetish btw>>101932141This one is actually me btw
>>101932529no I just love women mashallah
>>101932521wait for SD3 8B
>>101932544you are an absolute fucking maniacgen more sheep
>>101932529lower the guidance (comfy) or Distilled CFG Scale(forge)
>>101932544liar, you didn't made my post.
>>101932564no way thats not overlayed in post, right?
>>101932584You mean (We) didn't make our post, anon.
prompt: realistic beavis and butthead
>>101932587You underestimated the power of boomer prooompting>This is a digitally manipulated photograph featuring a group of african-american men and an anime character. The men are positioned in the background of the image, taking a group selfie while flipping gang signs. They are dressed in various durags, hoodies, tops, gold chains and shirts. >In the background, there is a an unmaintained ghetto area. The sky appears to be overcast, suggesting it is not a sunny day.>Superimposed on the left side of the image is a digital illustration of an anime character. This character has long, red hair tied into twin ponytails with large bows, and she has large, expressive orange eyes. She is wearing a school uniform consisting of a pink blazer with a white shirt and a red ribbon tied at the neck. She is also making a gang sign with her right hand.>The overall style of the image blends the realistic photograph with the vibrant, colorful, and stylized anime character, creating a playful and whimsical effect.
>>101932589i made the post, not the image.
>>101930774I'm training a facial cumshot Lora on a dataset of exactly half anime and half photorealistic images as we speak on CivitAI, samples look great so far. I just lead in on all the anime ones with `2d, anime` and on all the realistic ones with `photo, realistic` and the tagged them normally beyond that. Flux has no problem with either tag captioning or natural language captioning (or a mix of both) as far as I can tell.
>>101932611ermmmm, didn't ask
>>101932616>half photorealistic imagesyou mean photos? stop calling photographs "photorealistic"
>>101932624some of them were literal IRL porn photos, some of them were hi-res-fixed gens from realistic Pony models.
https://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/I just changed a feature on the Tonemap node so that the multiplier changes automatically when you change the cfg, so you won't have to change the both of them manually. You need to download the script again to make it work.
believe it or not, >>101932544 is also me.
>>101932617STOP IT, NOW! THAT'S LEWD
>>101932617They can take it further
>>101932606topfuckingkekboomer prompting to the fuckin moon
>>101932606where did the parent advisory sticker come from
>>101932646>some of them are gens…Is it a good idea to feed ai-produced pictures to ai?
>>101932681hoo buddy, you're gonna wanna sit down for this one; they've been training base image and text models like this for ages.
>>101932679I added> There is a "Parental Advisory Explicit Content" label on the bottom left of the album. The image is grainy, due to it being taken by a professional camera rather than a smartphone.to this prompt
Madison Beer Flux Lora is coming along nicely too (the training is 1024, just the samples are 512)
What is the local equivalent of NAI's "vibe transfer" feature?
>>101932606>>101932661>>101932703please stop sharing niggers, did you even gen them? i'm guessing it's just the anime girl part that's genned, at which point it's barely even a notch above ((1girl)).
>>101932719Lora?
>>101932723bait
>>101932681100% of the Pony ones were 1.5x hi-res fixed gens, so like 1.5x times more than the output resolution they'll be bucketed at. Quality is fine this way, where you run into trouble is using AI images that are the same resolution as your intended output res.
>>101932681ai generated images have no license
>>101932719some ip-adapter, there are several, google it
>>101932705Training on Kohya? I'm having no luck getting a usable output could you share your config?
>>101932723BLM
>>101932731you know what you're right i shouldn't have even responded to him - >>101932745 and there it is
>>101932752bait
war...war never changes
>>101932752It's over whitey. AI is the black man's playground.
>>101931372>Redownloaded the models and now it's fine.How is this related? Stop spreading misinformation.
>>101932778GO BACK
>>101932781even with that hoof, would
>>101932773how much test did they put her on?
>>101932796enough to get her a buttchin
>be me, literal retard codelet>need to make captions for 500 image dataset>get joy caption set up and working yey>tedious to input each image 1 by 1>spend literally 3 hours dancing with chatgpt to get it to re-write the app.py so that I can upload the images as a queue and have it spit out individual .txt files for each caption>one problem after another>it finally werksyeyI fucking love AI, it enables even retards like me to do thingsI can't seem to get it to make the .txt files spit out the same name as the images are named as (it did before but then it wasn't captioning them correctly), but since I have to go through and check each caption manually anyway this should be fineI'm happy. each day I am a step closer to making a billion flux loras
>>101932773her chin has seen some shit.>>101932661those dudes look like they mean business
>>101932744I'm just using Civit as always since they have it now, these are reflective of my "general concept" Flux settings:{ "engine": "kohya", "unetLR": 0.0005, "clipSkip": 1, "loraType": "lora", "keepTokens": 0, "networkDim": 16, "numRepeats": 1, "resolution": 1024, "lrScheduler": "cosine_with_restarts", "minSnrGamma": 5, "noiseOffset": 0.1, "targetSteps": 2048, "enableBucket": true, "networkAlpha": 16, "optimizerType": "AdamW8Bit", "textEncoderLR": 0, "maxTrainEpochs": 64, "shuffleCaption": false, "trainBatchSize": 4, "flipAugmentation": true, "lrSchedulerNumCycles": 3}you'd change the epochs based on image count to keep steps roughly around 2000 - 2500 or up to 3000 even if you want to spend the buzz.For the Madison Beer one everything is the same except the batch size is lower, 2 instead of 4 since that tends to give higher accuracy on single subject Loras where you don't need as much blending.Dim 16 / Alpha 16 produces Loras that are still a bit smaller than XL ones, around 150 MB, so roughly the same size as an SD 1.5 Lora at Dim 128 / Alpha 64 would come out. I haven't tried lower Dim than 16 with Flux.
Do dialects or accents show up well in text? If I wanted a french character to speak english with a french accent would the chatbot accommodate that by doing like "zee" instead of "the" or would it end the sentence "in a french accent"
how the ACTUAL FUCK DO I ACCESS COMFYSHIT'S SETTINGSthere used to be a cog icon on the queue sidebar (picrel) iirc but now it's vanished?also comfy used to run relatively fine a couple of days ago, but now it's slow as shit and literally unusable with flux for some reasonfirst it gives me an OOM error so i have to press "queue prompt" twice, and then it takes literal MINUTES to complete one (1) sampling step on a 3090what's changed?
>>101932792why would I go back to funnyjunk? 4chan is way more awesomer, all my friends think im a cool hacker when i talk about it
>>101932845tick extra options you silly dick
>>101932723my bad> This is a digitally manipulated photograph featuring a group of fat white nerdy american men and an anime character. The men are positioned in the background of the image, taking a group selfie while flipping peace signs and holding guns in the air. They are dressed in various tshirts, hoodies, vr headsets, funko pop figures and shirts. In the background, the sky appears to be overcast, suggesting it is not a sunny day. Superimposed on the left side of the image is a digital illustration of an shoju anime character. This character has long, red hair tied into twin ponytails with large bows, and she has large, expressive orange eyes. She is wearing a school uniform consisting of a pink blazer with a white shirt and a red ribbon tied at the neck. She is also making a gang sign with her right hand. The overall style of the image blends the realistic photograph with the vibrant, colorful, and stylized anime character, creating a playful and whimsical effect.
>>101932861extra options just reveals a slider for batch count and a checkbox for auto queuingi could've sworn there was a dedicated settings menu since i remember using it...though i guess there's a possibility it came from an extension i somehow no longer have, does comfyui not come with a menu by default?
>>101932839Thanks, I'm trying to get it working locally, I'll try using some of your settings in the gui
>>101932881was this supposed to be like a gotcha at me? because that is objectively funny.
you can always make time for fishing
>>101932904one can never gotcha a retard
>>101932861>>101932894nvm i'm fucking retarded, it's an extension that was disabling emojis kekstill not sure why it's so fucking slow now though
>>101932723>>101932881as a white man - sincerely, i finally feel represented, thank you. it's truly refreshing to see a diverse range of depictions that include individuals who look exactly like me.representation does matter. and your commitment to this inclusivity makes a significant difference. it fosters a sense of belonging and validation, one that my people sorely lack - and is deeply appreciated. thank you for your work in creating a more inclusive and representative digital space.all the best, anon.
>>101932606nice
>>101932828Yeah, I love using it for that kind of thing as well. Unironically I would not be using Linux today if I did not use AI to help me out with it and to write scripts for various stuff.
>>101932857literally no one talks like that, fuck off redditor lmao
>>101932927it was the word "fosters" that immediately tipped off this was generated by an LLMnice
>>101932934sliber ukranian
>>101932942FUCK
>>101932927i used this post as a prompt and got this
>>101932940what are you talking about? literally all my friends get real scared when i mention 4chan. it's literally the home of internet badasses xd so im kind of a big deal thanks to you /b/ros... haha
>>101932918>still not sure why it's so fucking slow now thoughIn my PC the UI itself takes up to 10% 3D processing lol
>>101932957>>101932907>>101932844>>101932812>>101932773+rep for war posting, got me back in line
>>101932975do you have a CPU with integrated GPU? you can run your DE off of that to gain some performance and VRAM back.
>>101932952you still get credit for making me lolnot your fault ive been fixated on llms since summer last year, i know all the slopisms.
>>101932897no problem. those settings might be hard to run locally keep in mind, depending on what you have, on Civit it doesn't really matter since they run the training on crazy enterprise hardware lol
>>101932975thing is it used to run fine with this exact same workflow a few days agothe only thing i did today was update one of the xlabs controlnets from v1 to v2 and iirc there's no difference in size between the twonow i'm consistently getting OOM errors and slow-as-fuck gensliterally unusable
>>101932911the joy that comes from implementing some really cool ideas and making a workflow for it
>>101933020No, just dGPU. It's happening after the new UI update.
How the fuck do you subclass a pytorch tensor to add some metadata without it being stripped by literally any operation jesus christ
>>101933035Yeah, I'm training at 512x512 which a lot of people have supposedly had success with on Flux
>>101932939it's really amazing, I can't imagine how many people don't realize the amount of shit they can do with just chatpgt alone. I've been trying to get friends/family to use it more but they're all so reluctant, no one understands how much it help with so many basic thingsit even understands if you want it to write you a custom comfy node. sometimes its stupid about it, but if you keep wrangling it, its gold
>>101933111> checkedi use it daily for work as a software engineer, and i always tell my friends i give it another 2 years before im easily replaceable by some jeet with a llm. trying to pivot into becoming just a pm so im ready for when everyone outsources
>>101933111Everyone at work uses it. We get it to write powershell and sql scripts for us. Saves us like 2-5 minutes each over doing it manually time but it really adds up.
>>101933137>>101933111>>101933151Imagine thinking what 2024 would be like 4 years ago. We're truly living in the future finally.
Where's the bake btw?
>>101933151>>101933137only a matter of time. wish everyone saw AI like this, artists even could use it to streamline their workflows and get new ideas of their own art - managed to get some artist friends using it for this purpose by training loras on their art for them, even. its truly the future
I like how she's holding the sword in this one. Just casual, playful. Another day at the park. What do you call this type of sword with an extremely long handle anyway? Is it even a thing?
>flux can't learn from booru tag style captionsI'm glad I didn't listen to the retards saying this. 15 epochs into training a flux lora on 2000 real photos with tags, no natural language captions at all, and it abso-fucking-lutely learns it, much faster than SDXL even. Maybe with these tiny 20 image datasets people are training celeb loras on it doesn't work, but I'm seeing excellent results.
>>101933192has anyone tried using both booru tags and boomer prompts in their dataset yet?I'm working on it but its gonna be maybe 2~3 days before I'm done since my dataset is large. I am a little worried it will learn that having booru tags then a boomer prompt is the 'correct way' to prompt the images, which isn't really what I want
>>101933181>>101933187yeah its absolutely absurd, i remember the original midjourny version and the garbage that made, but still blew my mind. it feels quite literally like "magic" to meim insanely excited about vr and ar as well, but way less innovation in that industry at the moment versus ai, but it will definitely leak into the headset area soo enough with ai offloaded rendering optimizations and suchplus, when i was young, i only dreamed of being able to just generate anything I wanted porn wise, and that came decades before I thought it was. the bleeding edge is soo sweet
>>101933111Well, at least for me, It could only help with problems that were already solved in a million different ways by other people. Maybe, I will have to train my own llm. LOL
>>101933221I don't think it should have both, it should randomly pick between using the tags or using the natural caption.that way it should learn it can be prompted with either mode
>>101933221Some model creators have said they do this, for example Lustify, off the top of my head. I think an ideal setup would be to have both a natural language caption and a set of tags for each image. 1/3 of the time train on the boomer prompt only, 1/3 just the tags, and 1/3 both. With a sufficiently powerful text encoder it should learn to be responsive to both types, or a combination.
>>101933221To prevent forgetting, for not just image models, usually what people do is mix in some old data with the new. So optimally some portion of your dataset will be boomer prompt, some will be tags, some will be both.
baker
Here>>101933401>>101933401>>101933401
>>101932778kys plebbit faganon
>>101933276>>101933269>>101933268thanks anons, I will test the various options and see how each turns out
>>101931837hisone to maso-tan?
