Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101972156>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>>>/g/sdg>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
Blessed thread of frenship
>>101975777jesus christ why is ai still so fucking bad... it's been 3 years and shit still looks like glossy slop
>>101975777>all of the retards training Lora’s and fine tunes on stable diffusion outputsLiterally why, you’re ruining the flux model with shitty deformed airbrushed AI slop
>>101975810>ZooeyYes.
>>101975824this, 100% this, I hate that AI imbreeding is being normalised this bad
>>101975808
i made it into the collage with that image? ok guess its good enough
What's the criteria for the making the collage
>>101975824anon you can't ruin this ugly piece of shit any further. this whole place is just retarded coomers and tinker trannies
>>101975847Why it always adds necklaces or chokers if you prompt for "goth"
>>101975824How else am I supposed to train my OC that I made with AI?
>>101975852It is good enough for /ldg/, lol.>>101975868Fried te.
>>101975888>896x512get a job
>>101975862just gen an image the collage maker wants to put in theresimple as
>>101975888Revert the size
>>101975897hahah I'm using fastflux.ai I can't use my computer right now
>>101975875these are flux pixel art? neat, i didn't realise. and i mean that in a positive way, i thought this was pony+lora, i'm guessing the occasional vectory outputs like >>101975890 are the model making mistakes when asked for pixel art?
>>101975922You can't "revert the size" (swap dimensions) with fastflux. It's just a tech demo.She's wearing a thick pair of grandma long stockings anyway.
ah man, I'm too much of a promptlet to make her part her thighs
>>101975955What's with this severe artifacting on her hair? The whole gen looks mangled...
>>101975955spread legslegs apart
>>101975955>she is spreading her legsnot rocket science
>>101975816This model has a default look that is very glossy in previous models you had to ask for it. They probably don't want it to be photorealistic out of the box. At least that's my hope. Otherwise it would mean that they have terrible taste.
>>101975998You think grok devs asked them to not make it photorealistic?
>>101975940Correct.
>>101975969that's the Flux grid seeping through
>>101976023This looks real
The keywords for different ages seem to be a combination of young/(nothing)/old and girl/female/woman.young girl: childyoung female: teenager, early 20sgirl/young woman: 30sold girl: 60sold woman/old female: 60s+The main take away right now is that "female" makes women more feminine, and never results in children. Also, there seems to be no reliable way to generate hebe and teenagers. There's an age gap from 10 to 20 year olds.>>101976074My new card <3
>>101976087>>101976023Use same seed until you get one with the hands fixed
>>101975907Norf FC spanish holiday
>>101975868>necklaces or chokers if you prompt for "goth"training data, but also the best part of goth girls (tattoos are the worst part)>>101975888i've never been able to make a skirt lift with panties showing>>101975955ime prompting for legs spread doesn't get you super good results unless you full on make them do the splits and even then flux isn't too good at it>>101976086>there seems to be no reliable way to generate hebe and teenagers. There's an age gap from 10 to 20 year olds.i've had success (like 30% consistency) with "young teen girl" or "tween girl", still often goes below 10 though but 11-14 is twice as easy to get as 14-17 I've found
whats worse is sometimes flux combines old and young features so you get uncanny anachronistic faces
when will Comfy stop unloading the model with every lora change?
>>101976128>tweenI'm not English native, and "tween" for me is something you did in Macromedia Flash. I didn't know it meant "preteen".>14-17That's were the weird gap seems to be. There has to be a keyword for it, but I don't want to research it non-locally because I'll end up on a list. My interest is purely curiosity, mind you.>>101976170They look very... Olympic? Like straight out of Asgard or something.
>>101976086Guess which one is this
>>101975969no idea. maybe because of the higher res>>101975978>>101975997no give. the finetunes can't come fast enough
>>101976182I don't know and that's annoying, it shouldn't work that way, you could simply keep in memory the original weights being modified by the lora and then bring them back for every change instead of destroying the whole model and recreating a new one
>>101976245loras affect the whole model
>>101976253no, it only affects a tony % of the weights, or else a lora would be the same size of the model, and you can see it's not the case, it's only 20-30 mb wheras the whole model is 23gb
>>101976245just reverse the merge using the lora weights
>>101976222Young girl? But that hair and other details means there's more elements to that prompt, so I don't know.The androgyny of children creeps me the fuck out. Creepy fuckers.
>>101976264I beg you go read how lora works before you make more a fool of yourself.
>>101976275that's funny because that's exactly the same reaction I got by reading your retarded take
trying to load this WF from here, manager can't find the "overridemodeldevice" node. can someone link me it?
>>101976295anon please, pull up the paper or an implementation and check your assumptions
>>101976205>There has to be a keyword for it,the keyword for it is "teenager" or "teen", which should by definition get you a 13-19 year old. there is literally no other term other than "jailbait" to describe that age range and i seriously doubt the caption model used for flux used that word>>101976205>They look very... Olympic? Like straight out of Asgard or something.it's probably the fact that the prompt is prompting for "armor" instead of making them have cybernetic bodies, since that's how joycaption captioned my reference image the goal is to make a bunch of cyberpunk doll-like girls for a music video>>101976222my bet is on "young female" because of the hair
>>101976298I'm guessing its this https://www.reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101976313now you have to explain to me how a lora can have all the weights while being 1000x lighter than the real model
>>101976264>or else a lora would be the same size of the model,I don't know myself whether it's technically viable to swap lora weights around without reloading the entire model, but this claim is not backed up by anything so I'm inclined to believe >>101976275 is likely correct.Out of interest, which UI currently allows you to swap lora weights without a model reload?
>>101976330that's what the rank size controlsplease, I bed you, go read an implementation
>>101976356*I beg youbut also I'll bed you if you don't do what I'm asking you to do
>>101976356I'll ask the question again, like a LLM maybe I'll get a good seed and a good answer, how many % of the model's weights a Lora is changing
>>101976341>Out of interest, which UI currently allows you to swap lora weights without a model reload?I think none? Forge does the same shit
>>101976379all of them, please, anon, google it, read the paper, read an implementation
>>101976379Why don't you go read how it's implemented instead?
>>101976318>the keyword for it is "teenager"In my experience, if you prompt "teenager", you get a 20-something year old, which is how teenagers are portrayed in TV and film, incidentally. There seems to be no way to prompt for actual awkward teenagers. They all look like 20+ Ukranian supermodels.
given how good flux is, give me one good reason why I shouldn't delete all my sd1.5, sdxl and, given that I don't do anime, my realistic pony models and loras?
>>101976397>>101976412>all of them,And... you losthttps://gitlostmurali.com/machine-learning/data-science/lora-qlora>his is the benefit of using LoRA. Just store the matrices into your disk, which would be maybe 1% of the original model weights
>>101976414Unless the answer is in using the CLIP text encoder and use good old SD1 tags. I have not tried that.
>>101976272>>101976318It's 'girl', but the entire prompt seems take into account as well. Like if the environment is a nightclub even with 'girls' it will gen teen+
>>101976397>>101976412he is almost certainly too stupid to understand and was either trolling and wasting time from the start, or has decided to start wasting your time now after feeling embarrassed about having posted something wrong online. please don't bother replying to him again <3
>>101976469and, you also lost >>101976429
>>101976379Don't bother. It doesn't matter that you're correct, others will believe whoever shouts loudest. He can't explain how https://huggingface.co/XLabs-AI/flux-RealismLora/tree/main?show_file_info=lora.safetensors which only contains lora weights for the double blocks, and only certain parts of the double blocks, affects the whole model. He's stuck on the rank aspect which does reduce the size compared to the full weights. Explain all you want, ask him to elaborate, it won't work.
Is everything alright in here?
>>101976417Flux is niggerlicious
>>101976505butthurt sai shills like usual
>>101976487I mean, you must be a special kind of retard to believe a Lora has the 12 billions weights inside of it while only being a 30 Mo file, but yeah, I'm done educating those retards, it's ok to be a low IQ, but not being a bad faith low IQ
>>101975843Cute and accurate rendering of a blessed thread of frenship
I will come back when AI can do this
>>101976579I guess that's a real picture right?
>>101976487he is not correct as shown by the sentence he cherry picked from the article here>>101976429, he doesn't understand how it works otherwise he wouldn't have quoted thatyou want to get technical, sure, loras don't touch all the weights but the loras all touch the same weights which is what matters to the discussionthe loras he sees on civitai all touch the same weights>>101976553see, he doesn't fucking get itanon, I beg you again, go read up on how lora worksyou can make it kilobytes large and still touch all parameters
>>101976462I see. Makes sense, since this model does not work with individual tags. It's more abstract.>>101976222What is the actual prompt here, if I may ask?>>101976579You can make a LoRA for something like that, with amateur pictures. SD1 can do it.
>>101976379>I'll ask the question again, like a LLM maybe I'll get a good seed and a good answer, how many % of the model's weights a Lora is changing>>101976397>all of them>>101976592>sure, loras don't touch all the weightsoops...
>>101976233Damn, what are you using for this? dev 50 steps?
>>101976649That looks like good old SD 1.5
>>101976609https://www.entrypointai.com/blog/lora-fine-tuning/lawl, unless you go for a 8,192+ Rank Lora (who does that?), you can see it in the 1% of total weights max
>>101976649kek, flux newfags
>>101976592What matters to the discussion is whether the entire model needs to be destroyed and reloaded. It does not because loras do not affect the entire model, loras are in fact configurable to only affect certain types of module and only certain layers.
Not that I'm upset with this view, but how do I get a from_below view without being able to booru tag prompt? have tried various permutations of "looking down at the viewer" and "worm's eye view" and "view from below"/"view from the ground". 1girl in her room shots are heavily biased towards shots from above, I guess, and I just need to have her in a throne or something.
>>101976649SD1.5
>>101976737low perspective
>>101976736And when you look at this board you can see it's not even affecting 1% of the weights when you go for a Lora Rank 512, so yeah, there's absolutely no point of destroying the model just because 1% of it got changed, the 1% that will be modified should stay in memory and be used to return the model back to its original form, that would be a way less retarded thing to do rather than doing this loading/unloading monkey retarded shit
>>101976684>>101976709>>101976744So why does flux not look as good then?
>>101976779SD1.5 is king when it comes to normal, ordinary looking picsFlux is impressive, but still in its early days imo
>>101976816SDXL is also very flexible. I think the secret sauce with SD1 was the wildcard dataset. We will never see a model like that again.
>>101976779loras, hires fix, cherrypicking, outright model finetunes toward specific goals, and base models being trained with somewhat "safe" datasets that make them worse at usually those same specific goals. the fair comparison is base sd1.5 -> sdxl -> flux, not some realism finetune that only does asian girls in egirl clothing taking selfies where the same splayed-fingers-around-a-phone block of picture is in nearly every gen. Not to disparage that anon's image at all, it's just that he's probably making it with a set of tools built towards creating exactly that, and from the resolution clearly has a big workflow set up around it boosting it up in size (I'd have assumed it was sdxl based to be honest)
>>101976737Damn>>101976751I've been wanting to know too
>>101976816>>101976832both SD15 and SDXL are RNG shitshows cause CLIP doesn't understand shit you tell it, ofc you can enforce what you want with a lore or finetune, but you don't have real creative freedom to prompt what you actually want
>>101976244That's a REALLY consistent pixel art image. Is that from Flux?
>>101976862>I'd have assumed it was sdxl based to be honestWe SD1 vramlets know it well. All the people SD can do are the same person with minor plastic surgery.>>101976897That's not true. It just works differently. If you have ever used NovelAI's v3 model you know how powerful a well curated tag dataset can be. In many ways I prefer it to Flux's natural language because it's more flexible and predictible.
>>101976832a big part of it is how loras can affect lighting. It is pretty much impossible to do stuff like this in XL or Flux. You can get dramatic lighting but it is way more limited compared to the contrast/lighting control loras in 1.5
can someone verify that loras work with Q4_0? I just want to test what I've trained
>>101976908Yeah.
>>101976609anon you're still not getting it>>101976708>>101976773that is not a % of weights touched, that is just a size comparisonlet me explain this to youyou have 5 parameters[0, 1, 2, 3, 4]you want to modify them by some deltayou could do 5 parameters[-1, 2, -3, 1, 5] giving you [-1, 3, -1, 4, 9]or you could do just one parameter applied to all of them[3] giving you [3, 4, 5, 6, 7]one has 5 parameters, that's 100% of the original weightsthe other has 1 parameter, that's 20% of the original weightsthis percentage is what you're seeing in that tableone is five times larger than the other, that's how one lora can be several gigabytes and another just 8 megabytesbut in this example BOTH ARE TOUCHING 100% OF THE PARAMETERS
>>101976779flux is cool
>>101976609>>>101976397>>all of them>>>101976592>>sure, loras don't touch all the weights>>101976945>100% OF THE PARAMETERS
>>101976751thanks, that's working a little better. it avoids the trap of sounding in isolation like it could be referring to the camera or to the subject like "from below" and "looking down" could, while not being as weird/rare as worm's eye view. not quite working but is at least focusing the bottom quarter of the image on legs with some foreshortening effect which is obviously along the lines of what i was asking it for. unfortunately all the obese dominatrixes in the dataset are making her displeasingly chunky now in 75% of gens.>>101976923they do on comfy and i believe worked on forge before comfy too>>101976914NAIv3 and Pony both have very strong adherence to tags so you don't even need to do the magic NAI did, although it's nice to not do the extra stupid things affecting Pony.
>>101976969yes
>>101976945That's a nice hypothetical situation that doesn't apply to the real world. In reality you have something like >double_blocks.0->19>single_blocks.0->38each containing>some_other_bullshit>more_bullshit>another_load_of_shitand the lora contains>double_blocks.0->19each containing>more_bullshitAs you can see, the lora does not affect the entire model.
>>101976945>100% OF THE PARAMETERSand 16 times the details
>>101976996>>101976773>And when you look at this board you can see it's not even affecting 1% of the weights I'm addressing this, anon, you misunderstood what that table was telling you>>101977008that's why I said "in this example", see above
>can train loras on the full size model on 12GB vram with -fp8, split_mode>can't load the full size model to gen with it on 12GB vram with fp8, putting vae on other gpu, model on main gpu, and clip on cpu, get immediate oomWhy?
Damn, Flux Q8 is giving me better results than the unquantized version
>>101976999>they do on comfy and i believe worked on forge before comfy toothank you, thought I saw someone say only q8 worked for loras for some reason. will DL baby quant for testing
>>101977036how much does your speed drop when you enable a lora?
>>101977020>you misunderstoodYou're the one who were claiming claiming that loras affect all the weights when it's a complete lie, don't forget that, you're full of shit anon >>101976609
>>101977025>Why?because the dev who coded the training shit is talented, Comfy isn't, simple as that
>>101977020Yes, we all know how matrix multiplication works, you've clearly grasped the concept of lora rank and how that makes the lora weights smaller. However, you are failing to understand that loras simply do not affect every module of every layer, that fact is why the entire model does not need to be destroyed and reloaded.
>>101977056and you misunderstood that table, as I showed youyou admit to misunderstanding that table and I'll admit to any mistakes I made, deal?
>>101977098can that retard not reference kohya's work or is he only capable of using other's code when its to spite illya?
>>101976999>they do on comfy and i believe worked on forge before comfy tooWhy are you lying? Forge just copy comfyui gguf nodes, they haven't contributed to making gguf work with flux at all.
>>101977136>you admit to misunderstanding that table and I'll admit to any mistakes I made, deal?I won't accept anything from you, you were grandstanding on your high horse on the lora subject by asking me to "read some documentations" as if you were some kind of expert when at the end you were saying some big stinky horseshit lies like "it affect all the weights", you're the one who needs to read the documentation anon, stop acting like a smartass when you can't even back it up with some true factual shit
>>101977153Comfy copied Forge's nf4 implementation
>>101977055Not much, from 1.5s/it to 1.7s/it
Does anyone know what the best site/resource to find the best (preferably local) ai tools? There'sAnAIForThat doesn't seem to search well and GitHub is alright but sometimes repos don't label it as ai, there's choice paralysis and it's hard to tell what the best tool is compared to others.
>>101977153I'm not lying, I literally did believe that.
>>101977205>inability to admit fault is a nasty trait, anon.and pretending to be an expert and grandstanding on others when at the end you were completely wrong is a much more nasty trait, anon
>>101977131>However, you are failing to understand that loras simply do not affect every module of every layeragain, that's why I said "in this example", in the example all parameters are modified to make it obviousI'm addressing something the other anon said that was wrong and I know you know what I'm talking about>>101977171inability to admit fault is a nasty trait, anon. you were looking for something supporting your argument and you completely misread it to make it fit. If you don't believe me ask the anon above. That table does not show what percentage of the weights are modified by lora rank.
>>101976923They work, but not so well with schnell.>>101976918>pretty much impossible to do stuff like this in FluxIt's still early to tell isn't it?
>>101977004bro, share your flux microbikini lora, the one on civitai sucks ass
I'm scare what have I done
>>101977223>inability to admit fault is a nasty trait, anon.and pretending to be an expert while grandstanding on others like a fucking faggot smartass when at the end you were completely wrong is a much more nasty trait, anon
>>101977240then be the bigger man and admit you misunderstood that article you linked, or do you believe you are correct in your interpretation of that table?cause you're not
>>101977163>i yearn for things that don't exist aaaah
>>101972120>>101972103Ok thx I'll try that when i get home
>>101977233You set guidance too low
As a long time A1111 user, who switched to ComfyUI to use flux since it was the "official" way, I just installed forge. Holy fuck it's so much better if all you're doing is basic shit. Like I get that if you have a complicated workflow that needs to run fully automated end-to-end, comfy is your only choice and it's fine for that. But the spaghetti node bullshit is fucking retarded if you just need to do basic text2img and img2img things, which is all I do. Forge just werks, is exactly like a1111, seems to have everything I need built in, it's just perfect. Shame the dev is a cocky autistic sperglord (tbf comfy is kinda the same way).
>>101977276I used a WF from here that had guidancepositive at 3.5 and negative at 7.0, what are the proper amounts?
>>101977255Did you admit your behavior was wrong by getting on your high horse and not actually backing up with some true statements? Doesn't seem like it
>hello is this thing on?
>>101976318if you leave out age descriptors for the subject and only reference age-associated events, locations, or relationships, the model will usually intuit and fill in where there's either no token or a "negative" token. for instance, an iconic teen bedroom with people in it or referring to classmates might create groups of people of similar ages. watching a pg-13 movie with a childhood friend, etc.I use this technique a lot to get around really strongly weighted tokens from training materials with a lack of variety. since you can't negate associated tokens in flux, you just have to avoid certain terms and hope the model can figure it out, which from what I've tested, it can in a lot of cases without LoRA. also, try using short sentences that chain or reinforce tangential aspects of the prompt instead of using weighting.
>>101976414why don't you guys use prompt editing to blend the words? it's always been the way to have control over this. [little girl|teen] or [little girl|little girl|teen] etc. should blend it into whatever age you're looking for. i don't know if this works in c*mfy though.
Schizoprompting gives you neat things sometimes>japanese 1990 kodak scan low definition grainy biomechanical dark lolita fashion>>101977306Ah, sorry, I don't do that, so I don't know now it affects the gen. The image you posted looks the way it comes out when you use a guidance of 1.0 with a regular workflow. That's why I said that.
>>101977314so you're not the bigger man then, you will not admit you were hastily googling for something supporting your point, you won't admit you latched on to a sentence that kinda looked like it proved you right but you actually did not know enough about the subject to realize it was talking about something else.you won't admit that
>>101977369So you don't want to admit your smartass attitude was wrong, I thought you were the bigger man than that
>>101977227in 1.5 you can train loras to affect contrast noise which allows for all those backlight/directional light control loras (that actually work). Based on my understanding it is impossible to do this in XL or Flux because you can't interact with that layer of the model.
>>101977364ah shit, I'll try the default comfy WF and see if that fixes it then. thanks for your help!
>>101977369>, you won't admit you latched on to a sentence that kinda looked like it proved you right but you actually did not know enough about the subject to realize it was talking about something else.>>101977136>I'll admit to any mistakes I made, deal?Oh no, he forgot to take his schizo pills again
>>101977304India?
what should i gen?
>>101977383nasty trait, anon>>101977400what's the issue there, anonie, you having problems understanding basic English?he admits he rushed for a gotcha and misunderstood what the article was talking about and I'll admit to any mistakes I made
>>101976940cleans up really nicely with extensions too https://imgsli.com/Mjg5MTI3/0/1 mostly whatever you're prompting seems to generate pixel patch sizes of 4x4 whereas the sakuemonq pony lora I use deliberately generates 8x8, making it hard to compare directly, but it seems to do almost as good of a job of trying to avoid generating overly detailed sections.
>>101977426>what's the issue there, anonie, you having problems understanding basic English?
>>101977421Two retards arguing
>>101977448not the nerd gif!
>>101977421pikacho being bbqed by mario
>>101977354this works well for the real world like bedrooms or school settings but its tougher in novel settings like the futuristic ones im going for>also, try using short sentencestough when im also trying to get a very particular look for the outfit but thank you anon, your instructions would be a useful prompt to enhance a prompt with an LLM for flux
>>101977466exactly picrel
>>101977466but make me the bigger retard
>>101977481It's fine, God already did.
>>101977485Ohohohoh
>>101977485God isn't real.
>>101977471throw this image into joycaption desu, i bet flux could generate this really well including the youtube page
>>101977364it's certainly something unique. weird reso tho
>>101977510Sure we can't physically see God, but we can physically see your fucking gunt (put it away)
>>101977510Then how did the universe got created? Can't make shit out of nothing right?
>>101977510says you
>>101977510god invented women, show some respect
>>101977530>that handQuality LoRAs
>>101977531without women we would still live in fucking Eden, FUCK
>>101977562damn
>>101977469What about the clip text encoder? That one works with tags, right? Doesn't that work with Flux?
>>101977518no you can't>>101977520okay then, god is that which was always existed and from what the universe was born... but you can't give it any more attributes now>>101977521yes>>101977531you're right, I do love my transsisters
Apparently this is Two Retards Fighting according to cyberrealistic(1.5)
The Gospel of Two Retards Fighting according to epicrealismXL_v8Kiss
>>101977590conjoined at the arm, the ultitard
>>101977561have one more
>>101977678the hand is rather distorted otherwise niec
Is there some insane finetune being cooked on flux or something? Did someone announce everything of that level?
>>101977700basically the opposite
>>101977680Thanks
>>101977700the gays are making one that can generate perfect dicks and gay sex positions
>>101977710feelsbad to know only the horse fuckers and the faggots are more motivated than the regular coomer simps on model training :(
>>101977721the solution is to broaden your tastes, why only get hard for 50% of the human population? there is a whole world out there.
>>101977697Ty
>>101977721it's a tough time to be a straight nigga
>>101977739box
Im using Q8 but it's still taking >22GB of VRAM, why is that? I can't load the controlnet like this
>>101977577>What about the clip text encoder? That one works with tags, right? Doesn't that work with Flux?these are great questions for someone who knows what they're doing. i just want consistent late-teen robot dollfaced cyberpunk girls>>101977721>horse fuckers and the faggots are more motivated than the regular coomerthis has always been true in AI and to a lesser extent in erotica etc, because if you have a fetish that you can't do in real life or find real content for (e.g. furry) you're much more motivated to create solutions
>>101977732can't change your sexuality anon, desu it would be fun to like men, I understand my type way more than fucking women that's for sure kekhttps://www.youtube.com/watch?v=S2ktZHBjkqE
>>101977749>late-teen robot dollfaced cyberpunk girlsLooks like you got it.
The Gospel of Two Retards Fighting according to realismengine_SDXLv30VAE (Bumped up base res and cfg scale 7)POV: You and your bro get into a fistfight over Halo 2 LAN circa 2003
>>101977766These guys faces reminded me of that one video of the nurse jacking off that retard tied to a wheelchair
>>101977746I can't
>>101974410>>101977631anyone got a catbox pls?
>>101977783>the nurse jacking off that retard tied to a wheelchairkek
>>101977783lmao
>two retards fighting
>>101977577>What about the clip text encoder? That one works with tags, right? Doesn't that work with Flux?not even Black Forest Anon that was here in the thread could answer me this question. In my observation its that it works with certain topics. Especially 2D content works okaish with tags in FLUX. Photorealism you want natural language prompts in t5. Also if you use text inputs for clip and t5 seperate, t5 mostly overides what you put into CLIP.
>>101977828desu the difference is so subtle I don't bother with it at all
>>101977762>Looks like you got it.thanks but a lot of the gens are too stylized, not photoreal enough to be consistently used in a piece of art together
>>101977739How the fuck is this not real
>>101977835well if she wasn't retarded before he showed up she sure is now
Looks like flux can learn weirder styles pretty well
>>101977846I like how these Flux gens are easy to inpaint.
>>101977917Beautiful
>>101977917woow, nice style
>civitai's downloader is down>file not found
>>101977941It's not Flux though
>>101977973so what is it
>>101977917lora?
>>101977973Better yet, no need for extra steps
It's over
We're so back
I hasn't even started
hoping we will see a pixelwave finetune of flux
>>101977978something more advanced
>>101978019Pixelwave sucks desu
>>101978005 so quick? >>101978009
>ONE GIRL>I DON'T KNOW WHY
>>101978046pixelwave is one of few models that gives unique outputs. other models will give you nearly the same images
>>101978063>TWO RETARDS>I DON'T KNOW WHY>IT DOESN'T EVEN MATTER HOW HARD I GEN>KEEP THAT IN MIND I DESIGNED THIS NODE SETUP TO EXPLAIN IN DUE TIME>ALL >I>DON'T KNOW
>>101955940this is more of what I was aiming for
anyone got flux dev to successfully convert to tensorrt?There seems to be some implied workingnesshttps://github.com/comfyanonymous/ComfyUI_TensorRTIf so, what are your specs and workflow nodes?
>>101977917Nice, the patterning on the spiral is interesting. Last time I remember seeing something similar was in Kamen Rider. https://youtu.be/FLFV-itlSzc?t=28
>>101977230you mean the one I used for pic related?
>>101977978SD1.5
>>101978149i'm asking what model
>>101978165Brav5
>>101978063Sorry SD can't do hands
>>101978124>anyone got flux dev to successfully convert to tensorrt?that would be fp16 right? you can't make quants out of tensorrt I think
so what's the best clip l? openai vit-l-14?
>>101978170which one specifically
>>101978146please more celebs in microbikinis
>>101978191yes
would there be any benefit in using this fp16 gguf variant of flux, compared to the regular fp16 .sft?https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-F16.gguf
>>101978170>brav5Holy shit the whole model is just that girl
>>101978205>yuri
>>101978201where 2 cop 336 version?
>>101978212no
>>101978222maybe it's this one?https://huggingface.co/openai/clip-vit-large-patch14-336
>>101978146yes pls
>>101978146yes or its inpainting?
>>101978232will comfy just load a .bin or .h5?
>>101978222you can't just swap the clip l for something different, the model won't understand the embeddingsyou need OpenAI's ViT-L-14
Not really feeling comfy today, bros
>>101978232found ithttps://github.com/openai/CLIP/issues/199
>>101978257What went wrong?
>>101978253if i dont fuck around then ill never fiind out
>>101978190I asked chatgpt4o it seemed quite happy to say that it can and included general suggestions of methods. I can't say if they are true or not.
>>101978277the finding out can be painful and with permanent injuries
>>101978277thishttps://www.youtube.com/watch?v=Of_8YG8b760
SD just gave me the perfect image and I beat it so hard my balls hurt
Is there any benefit of using the hires fix on flux?
>>101978257What's wrong with her arm?
>>101978295yes, use ultimate upscale to, pic related
>>101978323nothing, women have weird muscle attachments in the elbow, a degree of double jointedness is common in women.
>>101978193https://civitai.com/models/25494?modelVersionId=63786
>>101978384thanks
Flux.1 is getting pretty outdated at this point, Flux.2 when?
>>101978384kek'd at Stan Lee in the gallery
>>101978412right after SD3.1 drops
What forge version is best for pony and sdxl, or does it even matter?
>>101978412>Flux.2 when?when video flux drops + 8 months
>write some prompts in my air conditioned home office>refine until perfect>queue 100 images>go make some fresh gazpacho with 100% homegrown ingredients while they cook>come back to a perfectly targeted porn set with your particular fetishIs this what the Romans felt like just before their empire went to shit?
>>101978481yeah
>>101978412Black Forest Labs they were only created 4 months ago, so in 4 months they managed to create Flux.1It seems they were also working on the text to video model too. So I'm guess we will get video flux in 2 months and then 6 month after there might be a flux 2.
>>101978489Except they had actual human slaves to fuck instead of masturbating to blue light I guess.
>>101978492Lol, lmao even
>>101978477>>101978492you are all wrong, its just two more weeks as always
>>101978481This thing can't do teeth
>>101978498right
>>101978498A handful of them, the rest were poor fuckers working the field or deployed to bumfuck nowhere doing mandatory military service. Not that being deployed would be all bad, just fuck the local girls but there's a chance some German fucker would ambush and kill you.
>>101978323she's just a bit tired
>>101978510>This thing can't do teeththose look like very normal and real asian teeth anonthe real test is seeing if it can make kids with their front baby teeth missing, which i am not going to attempt for obvious reasons but also because flux can probably do it no problem
Let's roll...>>101978563>>101978563>>101978563
ive done some AB and cant see a difference on outputs between this clip_l linked by comfyhttps://huggingface.co/comfyanonymous/flux_text_encoders/tree/mainand the extra gigabyte and change on this onehttps://huggingface.co/openai/clip-vit-large-patch14/tree/main
>>101978237its the microbikini LoRa from civi tho
