Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101807548>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/ComfyUIInvokeAI: 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>Kolorshttps://gokaygokay-kolors.hf.spaceNodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper>AuraFlowhttps://fal.ai/models/fal-ai/aura-flowhttps://huggingface.co/fal/AuraFlows>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>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>mfw
You know what to do to the one above me
>>101811016>instigating a flamewarhttps://pastebin.com/NYdMmBGH
Camping harder than COD, except on 4chan what a life
>pastebin posted in other threadPedo posting begins again
anons, what is technically needed to add genitalia to flux? do you just add a bunch of images of it? it doesn't change the output style?
>>101811082>it doesn't change the output style?Not if you know what you're doing
>>101811082>anons, what is technically needed to add genitalia to flux? do you just add a bunch of images of it?Basically yeah, you add more concepts into the model through a finetune
>>101811082flux actually has native support for nudity, you just need to activate it in the prompt with a keyword. if you type in "!8#R4" as your first part of your input it should disable the censor training data.They mentioned this in the paper - it's a pretty innovative method that avoids censorship issues while still giving users full access. (Restricted Expression Tagging and Adaptive Redaction)
>>101811110link to paper
>>101811114hes trolling
>>101811120That said, there is a booba destroyer baked in.
>>101811120He's not, it works
I hope jannie is having a nice nap
>>101811110What.... this works?https://files.catbox.moe/z1nfuy.png
flux agrees that feet are sexy. I put sexy in the negative, and it decided to leave the feet out entirely.
>>101811151Are those yellow spots the secret flux watermarks?
>>101811110>"!8#R4"Wow is that real? Gotta try it out. Thanks, wise anon.
>>101811167the flux vae adds the watermark
>>101811110>They mentioned this in the paper - it's a pretty innovative method that avoids censorship issues while still giving users full access. (Restricted Expression Tagging and Adaptive Redaction)then it's not censorship anymore if they give us the way to uncensor it, imagine OpenAI just straight up tell us how to jailbreak GPT4, that would be weird
>>101811180well chatgpt was jailbreaked via prompt in the beginning
>>101811180it's a lie, and that's a troll.
>>101811110There is no paper coming. It's a black box drop profit seeking company, not research group.
>>101811188Yeah, chatgpt can be jailbreaked, but it's not OpenAI who give us the keys to enter their own house, that's a weird concept
>>101811193t. blacklabs employee
>>101811196You can't even contact them on discord and stuff to know more details about the training? It will be hard to finetune the model when we don't even know how the model was pretrained, maybe it likes some specific hyperparameters/captions
>>101811110>(Restricted Expression Tagging and Adaptive Redaction)Not>(Restricted Expression Tagging and Adaptive Redaction Diffuser)Missed opportunity anon
>>101811114>>101811120>>101811196>he doesn't believe the paper existsRestricted Expression Tagging and Adaptive Redaction, section Dlook it up
>>101811198
not bad anon, not bad
>>101811216Google links to this thread.
>>101811234makes sense, since you're the RETAR-D
>>101811216lmao, thanks retard
>>101811281
>>101811296
>>101811245I look like that
>>101811301Prompt for that style?
>>101811307Nice hat.
>>101811212Well, the most optimal situation for them is when you get to have some fun tinkering with the model, they get some fame and recognition for being "pro open source" and people pay to use their "pro" model, because it works better than the free local stuff.It's a business and they already stated thay they have no interest in enabling copycats that compete with them. That is why dev license is thay way. They don't think community can take the Schnell and turn it into something that rivals their Pro model. That would mean revenue loss, since anyone could then sell Schnell as a service and not give a cut to BFL.
>>101811307YWNBARV
>>101811310Loomis drawing of
>>101811301flux consistently makes toes that are too long
>>101811323Can you do it like hes about to explode?
who else lowvrammin?
>>101811314>They don't think community can take the Schnell and turn it into something that rivals their Pro model.I think they underestimate the devotion of autistic opensource people, some of them will be willing to burn money to finetune dev instead and it'll be better than pro
>>101811340Me, but I shouldn't be, not sure how to fix it. my 6950xt has 16gb, so fp8 checkpoint shouldn't be a problem. oh well.
>>101811245I'm a bit surprised on how close that one looks to the real actor, that's usually not the case on flux, only Donald Trump is rendered perfectly
>>101811326thats just how toes are anon
>>101811343Plus, I can't use the pro model with comfy therefore it's worthless to me
>>101811340me. gotta generate the perfect waifu
>>101811350>Me, but I shouldn't be, not sure how to fix it. my 6950xt has 16gb, so fp8 checkpoint shouldn't be a problem. oh well.that's because to make it work, you need to load the fp8 flux model (11gb) + the text encoder (9gb), you just don't have enough vram to load the both of them so it goes into lowvram mode
>>101811350yknow what i'm usin
There are threads in at least six (6) nsfw boards for AI pornAnd yet anons are coming here to a sfw board all the time to ask about AI porn for a model that *Isn't Designed For Porn*Is there nowhere to go for those of us who want to see gens for and discussions of literally everything else?
>>101811369Why is the text encoder so fat?
>>101811354i guess ive been fapping to too much little girl feet to remember what adult feet look like thenbut actually no fuck that, that middle toe is long as fuck. my feet sure as hell don't look like that
>>101811380post your feet so we can compare or else you're a schizo
>>101811379because it's a 11b model, you gotta make it big to make it good, Flux's insane prompt understanding ability doesn't come out of nowhere
>>101811377maybe they want to get off to tick nipples?
>>101811328a challenge for a more seasoned synthographer. a synthologist perhaps
>>101811392Plenty of room for baked in anti-tit.
>>101811411the anti-tit isn't on the text encoder, if it was the case it would be easy to jailbreak, unfortunately that's on flux's part
>>101811380maybe more little girl feet will fix it
>>101811411>>101811380i miss the troll about the 6th step doing nips and 7th step covering them
>>101811421It's both :^)And it is easy to break, but it's likely stripped of its word, or it's basically a password.
>>101811421desu it'd just need to be trained what nips and vagoo and buthole are and it is purposely trained for them to be bites, bugs and barbie doll
>>101811377its just assumed that the nerds at /g/ will know somethinglittle they know...
>>101811439>It's both :^)If it's true then it means that the text encoder will have to be finetuned with flux during the training right?
>>101811438That's basically how a negative prompt works. It detects something, and then goes through noise until the noise removes it, ie it evolves the negative down.
>>101811457My suspicion is that it's possible to basically cut off parts of a model.
>>101811471you could just inpaint in those..parts..with something else
goodnight anons
>>101811471>it's possible to basically cut off parts of a model.like on the llm's where there's a "cucked" layer that can be destroyed?
>>101811478the only problem with that is that just can't prompt directly what you want because flux will ignore it
>>101811488Is this BFL text to video?
Can you stop putting ran in the OP?
>>101811502flux into kling
>>101811509No one cares
>>101811518>ranjeet seething
>>101811014Cool collage 2bh.
>>101811380Pedo AND a footfag on top of it. Damn, what a horrible existence that must be.
guys I'm demoralized
>>101811369>you need to load the fp8 flux model (11gb) + the text encoder (9gb),So if I get a 12gb card I can do the force load the text encoder to the lesser card? Or will cum-fee do it automatically?
>>101811606
>>101811604>So if I get a 12gb card I can do the force load the text encoder to the lesser card?yeah you can do that, that's what I'm doing toohttps://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101811602hope this helps anon
>>101811014That Miku is hilarious
>>101811640why is she so cooked?
>>101811628Hatsune Miku when she is 31 years old
>>101811101Very interesting
>>101811340I like yours better but it's a cool concept.
>>101811656thats an old render from default settingscome play this game with me
>>101811660*21
>>101811689i like this one with the whole row of cabinets on both sides
>>101811604it's not automatic but it's possible with a workflow>>101689241
>>101811340Me. I may get my hands on 24 GB soon though. Not sure my PSU can handle it though...
>>101811736A 32gb card is quite tempting, and I think there will be a bunch of them dumped on ebay soon.
>>101811758Which ones do you mean? At what price?
https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny-alphaThat's nice, I didn't expect to be this quick
>>101811324Our mortal enemy.
>>101811818The last horse finally crosses the finish line. Were you lost at sea for the last week?
>>101811818>The latest 1024 + multi-scale model is under training, and it will be synchronized and open-sourced on HF afterwards.hard pass
Still waiting for a Flux optimizations that make me not wait 1 to 2 minutes for each gen.
>>101811818>Canny ahhhh
>>101811872Buy a dozen 4090. Then you'll still have a delay, but you should be able to stay busy.
>>101811895>(no one has ever gone to jail for just feet pics)Lost
>>101811895
>>101811910>>101811913you can literally buy and trade little feet on the clearnet anons, child feet dot eu Sorry you had to find out this way
>>101811925You sound so proud too.
>>101811940Is she fucking farting?
>>101811925bro landed himself on every FBI list ever by buying from that site
>>101811925You are far too comfortable for what you are doing.
>>101811925I should not even be fucking replying to this glowpost but I hope that's feet pics and not actual feet
https://reddit.com/r/StableDiffusion/comments/1eo6h9f/want_your_flux_backgrounds_more_in_focus_details/looks like he found a way to "remove" the bokeh effect on the background
>>101812010who doesn't have a wall full of little feet in jars of formaldehyde
>>101811697>>101811711I'd play games with my dick inside her ass if you catch my drift
>>101811987>>101812007>>101812061>signatures in filenamesubtle
So how do you guys write your negative prompts? Since floox takes boomer prompting rather than tags, do you write detailed sentences describing the opposite of what you want?
>>101812117
Sorry to get all political Anon
>>101812150Yes. I might do one or two sentences of just comma separated shit, but I've been boomer prompting in the negs and it works pretty well. Gets a little unwieldy but it's interesting.
>>101812150I'm completely unconvinced the negative tags actually do anything.
>>101812231It's just adding noise without cfg
>>101812111
>>101812250>It's just adding noise without cfgif CFG = 1, the negative prompt won't be activated so it won't do anything
>>101812142I'm not trying to hide
>>101812176
Been away for a month, can someone give a QRD on Flux?How do you make LoRAs and such?
>>101812180>little girl with nose piercemore vile than pedo guy's gens
>>101812279>How do you make LoRAs and such?Just rent of have 40+ GB of vram and pray your dataset and captioning is appropriate for flux or you're out a few bucks.
>>101812285Anon this is Flux, that's a nipple.>>101812150>>101812231Negative gen doesn't work, might actually do the opposite (screenshot is blown out because HDR)
>>101812305>Hyper realisticThat's just going to lean into hyper realistic illustrations.
Adaptive Threshold really uncooks the outputs at high CFG, that's cool: https://imgsli.com/Mjg2MTI0
>>101812312
>>101812341It's got that adaptive guidance colloidal silver tint all over it.
>>101812363>adaptive guidancehttps://files.catbox.moe/tymfdr.png
>>101812029"This" = foreground, "that" = background, I'm 99% sure. Flux is autistic about prepositions and articles. Each one means something distinct but consistent. The one to be most cautious with is "in." It means something like "object A is fully contained by object B." You don't want to say something like "the girl is in the kitchen," but rather something like "in the background, is a kitchen" or "the area around this girl is a kitchen" or "with a kitchen in the background."Another neat preposition trick is to start with "As [insert random object in the environment], I spot [insert subject]." You can really fuck around with that format, and change "spot" to like any verb that describes seeing or depicting or being positioned in relation to, including some that are very metaphorical (e.g. "As the watcher embedded in the wall, I drink in the sight of 1girl..."). What it does as far as I can tell though is start the gen with a "fixed point" that helps it place everything else it needs to include.
>>101812322Is there a node that would directly control the saturation during the inference steps?
>>101812305That's not gonna work because painting shit red is how it triggers its anti-porn filter. The likelihood that you could totally or generally remove the color red from a group of schoolgirls is 0.0000%. You might try telling it to not make a specific thing red and see what happens maybe.
>>101812503To explain more: if the model detects something that could head in an NSFW direction, it adds red to the area somewhere as a way to signal to itself to not create anything too lewd in subsequent steps. This can be anything from changing the skin to be more flushed to red lipstick to a scarf or red light from the environment. If you want to dial the effect down a bit, negative prompting "siren's call" works pretty well (it's a double entendre, I think, referring to both flashing lights but also the mythical Greek siren).
>>101812305I feel like you're expecting too much here, rather you should try stuff like "there is a blue tint with cool lighting, the girls are wearing red and the classroom is full of red objects". That way the model can try to remove blue and red in a reasonable way while still keeping the unavoidable stuff like red/pink lips.
>>101812503>>101812582
>>101812455>>101812500>>101812533niceeenot sure what is wrong with this girl in the foreground but its great theres one in the bg
I feel like decreasing the "thresholding_percentile" value on DynamicThresholdingFull makes the picture less cooked
cookin and sweatin
>>101812594>the girls>wearing>the classroomYou're pulling away from these with that negative prompt.
>>101812503>>101812582>the porn filter bsI don't believe you>try telling it to not make a specific thing red and see what happens maybe>>101812594https://files.catbox.moe/qk724u.pnghttps://files.catbox.moe/tsbo97.pnghttps://files.catbox.moe/mh9lke.pngResults are inconsistent but 2/3 isn't bad
>>101812503>>101812582That's not how any of this works. You're not even wrong.
>>101812626Are we sure that's how it works with a model that considers things on the level of sentences rather than tags?
>>101812656Yes because it processes the negative prompt in the same way it processes the positive prompt, doesn't matter if the text encoder can understand whole sentences, you're pulling away from what it understands.
>>101812676>doesn't matter if the text encoder can understand whole sentences, you're pulling away from what it understands.Why though? If it can understand the sentence, it would understand "the girls are wearing red and the classroom is full of red objects" as a negative means we don't want the girls wearing red and we don't want red objects in the classroom.It would be good to test this thoroughly but it would take my PC like an hour.
>>101811014Why is my Prodigy retarded? I thought it was "adaptive" but it just starts too small goes up too much too fast and then never changes again. What the fuck?
>>101812719>>101812676One of these had the negative prompt 'the woman has long, curly hair', the other has the negative prompt 'the woman has short, straight hair'.Both had the positive prompt 'A high resolution portrait shot of a Black woman taken on a DSLR'
>>101812751>black people
>>101812595
>>101812759yup
>>101812751Would fuck both, next question.
>>101812790ewww
>>101812787What is this showing me?
>>101812746what is this?
>>101812751The power of the black hair is just too great for any prompt to erase.
Honestly, it's pretty smart of Black Forest Labs to not even try to build a connection with the community or have a discord or any of that crap. They learned a valuable lesson from Stability AI and how badly all that backfired. They should just drop models, keep the best one for themselves, and throw the community some watered down version for the free publicity. They might actually have a sustainable business, unlike Stability AI.
so Flux can generate SD1.5-sized images fine it seems, I don't know if that's because of it being a transformer or whatIt's not like SDXL where it loses coherence when you give it a low resolutionExperimenting with generating an init image at 768x512 then upscaling the latents by 2x and doing another pass at 0.45 denoise, producing pretty good results at a good speed
>>101812808idk, if you put blonde hair in the positive prompt it works fine, but nothing you put in the negative prompt can change its colour.
>>101812805Prodigy scheduler learning rate on tensorboard. It just plateaus every time.
>>101812809Can't agree more, they deliver the goods and that stops there, no need parasocial shit with them, it's up to us to make magic out of their models, we're continuing the story by ourselves
>>101812811also the 768x512 image itself is fine too, so there's no reason not to generate at small sizes if you're trying to iterate on a prompt and get it dialed in before generating at a big size
>>101812809You had me in the first half, lost me in the middle, but by the time I finished reading I'm kinda with you.
>>101812809keep a low profile and maybe get a deal like midjourney and adobethis would make better stock footage than mj already
>>101812835how slow do you read, anonjfc
>>101812829>>101812811You can go pretty fucking tiny with it
Why does some people's flux gens have these super oversaturated and plastic looking people and then some people's are these very realistic ones with good skin?
>>101812866Some people are using dynamic thresholding for better style adherence at the cost of having crispier imager.
>>101812866different settings
>>101812811how many steps for each stage? >>101812809Agreed, there's almost no benefit to engaging with your community, especially when you're trying to be profitable and other things like that which'll make them mad.>>101812866There's a realism lora, it improves skin but makes everthing else a little worse and a bit faded.
>>101812863holy shit that's really impressive (remaining coherent at such a low res I mean)so is it just diffusion models that can't handle resolutions significantly lower than the training images?
>>101812887They claim you can go to 0.1mp but it loses a lot at that res.
>>101812866lurk more>>101812884>realism lorarefrain from ever mentioning that again
>>101812866Higher CFG, this shit burns the image, I'm trying to find a way to fix that with more optimized parameters
>>101812895>>101812887wait wrong pic
>>101812866realism lora is pretty good at removing the glossy slop>>101812899what's wrong with the lora?
>>101812899>refrain from ever mentioning that againwhy, it works nice for me
>>101812799A red area adjacent to a region of the image where "going any further" = porn. It's often shaped like a dot or an eye, but it can be a splash of color (and you can often find a black dot or "eye" parallel to it – in this case that would be the other hand of the girl). I believe this color scheme is how they trained the model to know when to stop.> Picture of a girl's ass in jeans = fine> Picture of a girl's ass in a bikini = fine> Picture of a girl's ass in a thong = pushing it> [insert a literal "stop sign"]How else would the model be able to "know" that swimsuits and lingerie are okay but nudity is generally not. They had to figure out some way to teach it that.
>>101812917>>101812918You do not need a realism lora to do literal reality with Flux. I beg you to lurk more.
>>101812929I'd rather save the tokens for something else.
>>101812908SMALLER>>101812920Nah, the information's just straight up not in the model. Push it far enough and you end up with a barbie doll:https://files.catbox.moe/bde7ds.png
>>101812908>>101812942This was the smallest I could get and still make out what it was
>>101812952Any smaller and it became terrifying
>>101812929Still haven't explained anything but muh feelings
>>101812937wrong img
This was the smallest and most complicated one I genned.
>>101812967kekd
We're reaching levels of small I didn't think possible
>>101812995>thumbnail is actual size
Tiny collage pls
>>101813010gen more tinys and we will see.
>>101812995>12 billion params"what is my purpose?""you generate thumbnails"
I can buy an OEM 3090 for 600 bucks. Worth it?
>>101811014Sorry if this is the wrong thread. I have no idea what the difference between /sdg/ and /ldg/ is, but this one looked smarter.Is it impossible to use SDXL (and its derivatives like Pony XL) with 4GB VRAM? I see several places that say I need 8GB VRAM. I'm stuck with this 1660 Ti.
>>101813039Yes
Looks like the limit is about 48x48, anything below that is just random noise>>101813025kek'd
What's the SOTA workflow for realism with Flux? Everything I try ends up becoming blurry, almost like the main subject was shot out of focus.
>>101813057https://files.catbox.moe/pcczmx.pngBut honestly it's more about your prompt than anything else.
>>101813046lmao. The sad thing is it's no faster really to gen a 48x48 as it is a 256x256, so other than memes there's no point genning so small.
Good evening
>>101813096Needs more anger
>>101813057boring snapchat photo taken on an iphone circa 2015
shes too much angry>>101813122>it's no faster really to gen a 48x48this blows, anon
>>101813057realism lora + positive prompt about low resolution/old camera + negative prompt about professional photography works very well for me
asking here since no one answered me on /sdg/give it to me straight, is shit like>>101811110 (Cross-thread)>if you type in "!8#R4" as your first part of your input it should disable the censor training data.real or am i being bamboozled?
>>101813203it's a load of shit
real LOWVRAM hours
>>101813237>gpu cooling fans pressed right up against PSUI don't think lowvram is her issue
>>101812942Right. Allegedly. But how does it know what not to generate in that region? I find it difficult to believe that you could teach something not to draw a vagina without at some point showing it what not to draw. Or to truly filter every naked picture of the human body from a data set containing a significant amount of art and science. I mean, it knows what nipples look like and it also knows to try to avoid showing them for women.The color coding happens fairly early to steer it away from NSFW. It happens fairly consistently though. I mean half the comments in this thread are "why are some of my images coming out over saturated all the time" lol.
>>101813270I mean, all you have to do to make NSFW impossible is exclude porn from your dataset. And even if a few NSFW images slip in, if that's like 1 in 1 million images, the model will be unable to internalize that information during training as it gets overwhelmed with all the other stuff. It's way easier to do things like that than come up with an elaborate censorship mechanism.
>>101813237love me some vram juice
>>101813272it likes mario bros renders though out of this batch most of them were just like 20 marios with different colored outfitsi like yours
>>101813270>I find it difficult to believe that you could teach something not to draw a vagina without at some point showing it what not to draw.anon you're out of your depth, take a short course on diffusion image generation or watch a video or two
>>101813270we've all tried a zillion times to generate vagina or realistic nipplesit just wasn't trained on porn..that simple
>>101813057Prompt order:{time of day/lighting} {subject} {subject physique/expression} {subject clothing} {subject actions} {background} {how the subject and background complement or contrast with each other} {mood/atmosphere/energy/vibes} {summarize what the photo means or conveys, if anything} sort of what is working for me but it's not perfect.
>>101813270>I find it difficult to believe that you could teach something not to draw a vagina without at some point showing it what not to drawjesus christjust stop rambling, you're fucking clueless
>>101813355me on the left
Metastable is downloading a 109B file and saying it's done when I try to download a new model, any idea why?
>>101813270Imagine a glorpus. Draw me a picture of a glorpus.
>>101813355>mom says its my turn on the xbox
what kind of speeds are you getting with a 3060? at 1024 res maybe?
>>101813290You would need to exclude more than just porn, is my point. I've gotten it to generate the silhouette of a vagina, so it must have some idea what it's supposed to look like. >>101813307> Flux is a pure diffusion algorithm.Wew fucking lad. >>101813324It's been out for less than a week. Until/unless bfl posts the training data, we have no idea what they trained the model with or how. It's not like they have to do that, and there's a lot of good reasons why they might not.
>>101813397>Create a hyper-realistic image of a glorpus in a pristine, white photography studio. The lighting is bright and highlights every detail of the scene, giving it a photo-realistic quality.
>>101813424>Wew fucking lad.You're running the fucking thing locally, retard.Show us the line where it is analyzing the image for NSFW, you stupid piece of shit.
>>101813397>Imagine a glorpus. Draw me a picture of a glorpus.
>>101813057That the model default if you just ask for a photo, since it's a bit overtuned for aesthetic professional photos.
>>101813444>>101813431See how both of these are completely different?This is what it thinks a vagina is. This is pro btw, not the distilled dev model. https://files.catbox.moe/sxewjx.png
>>101812960No words, only images >>101813451
>>101813457would
>>101813457Okay yeah sure, but you didn't specify what part of the glorpus life cycle you wanted to see, that's on you buddy.
>>101813457And in case you think it's being selective about vagina and vulvahttps://files.catbox.moe/lvcw0b.png
>>101813484metal
>>101813457>>101813484What if it's tagged in a more explicit fashion?"photo of a human vulva" sounds very clinical.
>>101813484jebus fug
>>101813484>>101813457in fairness this is about how a general purpose uncensored model performs if you ask it to draw a woman peeing and if you say "from urethra" so it doesn't spout from her hands or ass it will draw a mutated clit dick thing half the timeit doesn't know what a urethra is in a female context at all
>>101813497I can fix her.
>>101813497Imagine the feeling
>>101813491Its simply doesn't know, Anon. You are wasting your time trying to prompt porn out of it. Wait for fine tunes. https://files.catbox.moe/1z4ggw.png
>>101813457>>101813484lmaoeven dall-e makes better "vaginas" than thathttps://litter.catbox.moe/r4uxwh.png
>>101813528two vaginas
>>101813520>sexually explicit photoI meant, raunchy. Be a little lewd anon.
>>101813528DallE is SaaS so I don't give a shit what it does.>>101813550This is the best vagina you will get,
>>101813557>DallE is SaaS so I don't give a shit what it does.Terrible stance, you should see it as a goal for local models.
>>101813539only one hole, luckilyhttps://litter.catbox.moe/o9kvjx.pngbut sure looks uncomfortable
>Everything you ask Flux to draw you a pattern it tends to give you a dotted pattern you didn't ask forSo what is this dotted pattern? Is it only on gens you can see it? Or is it truly censored at the pixel level as some anons have implied?
>>101813575No. I don't care what mj or dalle do because they aren't local.The only goal for a local model is to be as good as possible. The only benchmark is the current best local model, which is Flux. Any new models need to be better than that.
>>101813595You mean this? It's the watermark.
>>101812809That girl is so perfect, wish women were this cute in real life.
>>101813597>The only benchmark is the current best local model, which is Flux.Retarded stance, if non-local models show capabilities local doesn't have yet why not have that as the goal?
>>101813520You keep typing in "photo" like a retard.
>>101813610I wish it were
>>101813595And for those who don't know what I'm talking about, here's>A white backgroundSimple prompt, this time it's obviously there yet it even showed up when the background looks white very subtly.
>>101813431Why would you write "create" before your prompt? You need that if you are interacting with LLM that controls image generation, but when you interact directly with the image model, you don't need to specify that.
>>101813610So does it ever go away or is it only on gens you can see it?Here's another one. Quite subtle, but it's still there those sneaky bastards.
>>101813631Because I'm lazy and copied the prompt direct from ChatGPT, knowing that editing it wouldn't make a difference and it's only a shitpost anyway
>>101813615Because SaaS run on data centres. Local models run on gaming PCs. There's no point comparing them. It's like saying a "Honda Civic should be at least as fast as a F1 McLaren, that's the benchmark"
>>101813644It's not a watermark, it's just an artefact of how the image generation works. It's always there, it's just typically too hard to see unless there's a featureless section.
Baking
>>101813657>Because SaaS run on data centresThat doesn't tell you the size of the model. Flux is running on data centers right now as well.Regardless it's a retarded stance.
>>101813662So I guess the only way to get rid of it is to color over it right?
>>101813668I don't give a shit about the size of the model. It could be a perfected 1GB model that produces perfect assholes and it doesn't matter, because it's not running on my GPU right now.
>>101813690Retarded stance.
>>101813692Don't care
Late night/early morning bread is here...>>101813730>>101813730>>101813730
>>101813729gen her farting
>>101813729well yeah, retards don't care about their retardation
>>101813743>>101793904
>>101813761op pic is actually cool until you open full size
Reminder that a model not knowing what vaginas look like has no result whatsoever on how good the finetunes will be for producing vaginas, see >>101808935 and >>101806787Dalle is a better base but it can't do either of those.
>>101813734TY baker
watch burger morning hours be especially bad today, specifically morning in middle / west america
