Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101986261>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
>>101988457thank you baker
>>101988388Oh no (((they))) got him too
>>101988457Can anyone explain why do all the AI models have this blurry, smudged look? It's like there has been no advancements made during last few years and people are stuck with the same tech. I've been looking at last several threads' """highlighs""" and it really doesn't look good.
>>101988505thats weirdly cool
>>101988500dataset issue
>>101988505This is amazing, catbox?
>>101988477psure your image selection would've been better.>>101988479she looks seriously concerned about the well-being of her puppets
>>101988500It's stalling thanks to those faggots crying over their furry porn drawings being "stolen"
>>101988532>>101988519Made on forge. Used dream lora and ps1 lora at 1 strength. this is the dream lora:https://civitai.com/models/663582/dreamy-floating-flux-loraPrompt:The bi-spectacled people are hysterical. Nothing here is real, as if it's a simulation. The landscape is beautiful, yet empty - as if a hollow mockery. Dragging something, what is it? Who are they? I hope they can't see me...but I have a feeling they can. Why are they smiling...
>>101988549>psure your image selection would've been better.Nope it was going to be only this one
>>101988532>>101988505samefag
>>101988549>she looks seriously concerned about the well-being of her puppetsthat was "a photorealistic flumpuszuch with its babies, inspired by HR Giger and Junji Ito"and yea seems to be caring
>>101988500>why do AI generated images all look like they are AI generated images?Because they are AI generated images.
>>101988573gay fag
>>101988563Awesome, thanks
>>101988563>Used dream lora and ps1 lora at 1 strength.sd loras?I've never even used a lora.
>>101988570well..>>101988578flux can do these things really well.
>>101988605Forgot to mention cfg 6 and Dynamic thresh, Euler Sgm Uniform 30 steps
>>101988611is that some Deborah A.W.?
>>101988591So this tech is ultimately unviable for anything except for shitposting and masturbation aid?
>flux
>>101988500>make the default look like stylized digital art >all gens by the masses immediately recognizable as sloppa>people leave you alone and don't accuse you of breaking the fabric of society or whatever
>>101988636yeah. finally did my lora accumulation trip. easy to work with and she's nice. angourie rice here is a bit of a bitch to dial in.
>>101988652nope, I see more and more content creators use it for videos and commercial media use it for illustrations (just be sure to say you are using it)
>>101988652currently? yes. but flux doesn't really have any good finetunes yet so we can't say for sure if it's salvageable or not
Blessed thread of frenship
https://civitai.com/models/665628/alex-jones-lora?modelVersionId=744933https://civitai.com/models/661908/wojak?modelVersionId=740736uh oh
>>101988500the companies that make models don't want to get in trouble so they try to give it an obvious "ai look" like dalle does
>>101988727No.
>>101988677
>>101988736yes, OpenAI can make insanely realistic shit if they really wanted, look at Sora it's their best example https://www.youtube.com/watch?v=h37A4zocIFg
>>101988736why no?
Does facedetailer work with Flux? It should, right?I normally use it to do face swaps
>>101988748>>101988745>the companies that make models don't want to get in trouble so they try to give it an obvious "ai look"no, dall-e being an example does not mean they all are doing that
>>101988500>no advancements made during last few years and people are stuck with the same tech.yes and no.flux has better text encoder, but the model is still based on 2D diffusion model.anyway 3D is the future. (text to 3D+animation, image to 3D+animation, ...)
>>101988744Kek
>>101988652you can do whatever the hell you want with stable diffusion. endless possibilities. >>101988724seen them lol>>101988749in comfy, yes. flux frighteningly good at inpainting
>>101988764okay, give me an example of a model that doesn't create ai slop.
>>101988764again, they can make way better quality pictures than dalle, they showed they can do better than that with Sora, yet they decided to stick with this plastic looking humans dalle outputs
>>101988775nijijourney
>>101988775ERRRRRRM They just don't. Okay chud???Kekpitalism always works
>>101988786yes, the one model that doesn't do realism, curious.
>>101988802can you upgrade her fingers
>>101988799you asked for non-ai slop outputs, not "realistic non-ai slop" outputs, you had to be more specific than that anon
>>101988799nothing weird about that, getting over uncanny valley with realism is much harder to achieve unless you have really low standards like a horny jeet
>>1019887653D is my Achilles heel. Not the sculpting. I love sculpting, and I'm pretty good at it imo, but retopology always ruins everything I do because I am shit at it.If AI can even just do that for me I will be happy. I have tried all the automatic retopo plugins and they are all shit.
>>101988724kek
>>101988814>>101988818all models capable of realism and "dangerous" content look slopped. models that aren't capable of that don't look slopped. why?
well it knows NIKE>>101988814just ignore it. it has its mind already made up. waste of time anon. bait
>>101988842>>>/x/
>>101988775Is it AI slop because they are purposefully making it worse or is it not yet perfect because they haven't figured out the perfect training method or there isn't enough compute yet.Going outside and not seeing any black swans doesn't mean all the black ones are being painted white.>>101988776Again, I wasn't talking about dall-e. Not that dall-e was perfect before the nerfs, it had obvious AI issues too, but less so than anything else at the time.And again, OpenAI nerfing dall-e does not imply all AI companies are nerfing their image models.
>>101988842did you completely ignore the uncanny valley argumentrealism looks like slop because it's much harder to replicate compared to much more simplistic anime style
>>101988808Sure.
>>101988842>models capable of realism content look sloppedbecause you have decades of experience looking at reality, dipshit
>>101988860>And again, OpenAI nerfing dall-e does not imply all AI companies are nerfing their image models.OpenAI and SAI do it, that's already a large % of the total relevant companies
>>101988881>SAI do itif you mean they filter any sort of possibly NSFW content that's a completely separate issue to what is being discussed
>>101988868flux's anime capabilities also look pretty slopped. if you're a company making a model capable of realistic content i don't see why you wouldn't intentionally nerf it to avoid bad press and regulation. openai does it, im sure black forest labs does it too.
>>101988914Nerfing sounds a lot like they are finetuning.
>>101988802Can you make him taking home a cunnybot
>>101988892>separateRemoving NFSW and making realistic AI slop pictures are different methods for the same goal, having the least controversy possible
>>101988500Found a trick to reduce the blur in schnell, but the ai slop flux look is still there
>>101988892>if you mean they filter any sort of possibly NSFW contentthat's a nerf, like you just called it, dunno how you decided to argue with that>>101988860>not imply all AI companies are nerfing their image models.>nerfingSD3M can't even lie humans on the grass, if that's not a nerf I don't know what to call it
>>101988938Slightly less blur
can't even load 2 loras at the same time without having ComfyUi shitting itself and crashing my PC, what a piece of shit of a sogtware
>>101988457Until this day, this output fills me with much joy. I wish I could replicate it with the prompt but there isn't a model for it. It was made using Bing. U_U
>>101988974of course, Flux is also censored
>>101988968Flux will can't get the details in the background like dalle
>>101988953still pretty fried>>101988959skill issue>>101988974dude you are on a blue board, and this is deep fried shite
>>101988914with loras it looks fine and it's only beginning. they probably didn't include enough anime and/or trained it on synthetic slop
>>101988959Sounds like you need more VRAM
>>101988983>>101988974The flux titty situation is a curious one. I belive we will uncuck it.
>>101988992it's already insane we are willing to compare a base model (flux) with one of the best API models that exist so far, with great finetunes we'll be eating really good
>>101988999It's not fried, it's schnell, that's the default look
>>101988999FALSE. Titties are natural, if God made titties who are you to clothe them?
>>101988939>that's a nerf, like you just called itim nta
>>101989013>it's schnell, that's the default lookyikes... I'm glad I never tried this shit in the first place
>>101988974Interestingly, nobody has made a seethrough lingerie LoRA yet. Could get a shit load of free buzz on Civit. How many images roughly do you need to train a concept? Only ever trained characters.
>>101988773Downloading Crystal Clear XL Prime. Basic advice, coming from Flux.
>>101989004if the 5090 doesn't come with more than 28GB then I at least want one of those spatulas that he seems to have a lot of.
>>101988831
>>101989012>>101989017>>101989007It either adds some shit like that or chains or the bra ends exactly on the nipple, if they made it for Grok I think they have some filter to censor NSFW in it
>>101988992Thanks. I'll look into it.All I know is that the prompter sort of used "oil painting cow drinking milk" as prompt. Hopefully Flux can get the rest.
>>101988959how much vram do you have, and what quant are you using?
>>101989033What the fuck I never noticed this. Why does he have so many? What could he need them for? We all know this nigger doesn't cook for himself.
>>101988939>that's a nerf, like you just called it, dunno how you decided to argue with thatBecause it is a separate issue, dipshit.Companies avoiding being associated with pornographic content is an old tale, it has fuck all to do with the issue.Again, if you go outside and see no black swans it doesn't mean OpenAI is going around painting the swans white. The non-existence of black swans doesn't mean it is OpenAI's doing.There are reasons models aren't perfect other than "they can do it but they nerf the models", fucking idiot.>>101988939>SD3M can't even lie humans on the grass, if that's not a nerf I don't know what to call itSD3 2B is a badly/undertrained trained piece of shit, do you think they WANTED for it to be that bad?>>101988974they nuked nudity from the data set on top of their caption model probably doing a shit job describing nudity if any slipped through
>>101989048it's not a vram overflow, I have 24gb of vram and I'm using Q8_0, like it just crash and my gpu makes a small peak at 100%, something wrong when you go for 2 loras on ComfyUi for whatever reason
>>101989038They didn't make it for Grok, they just licensed it out to them after the fact.
>>101988974>>101989007its not curious at all, nipples are censored in the dataset, genitalia are just not present in the dataset, it does not censor internally, they did that in the dataset. Therefore flux just doesn't know em at all, but look at civitai flux loras, you just have to relearn flux these concepts, atm all atempts are mid, but eventually it will work out.
>>101989062>if you go outside and see no black swans it doesn't mean OpenAI is going around painting the swans white. The non-existence of black swans doesn't mean it is OpenAI's doing.wrong! and if you hear hooves it's almost definitely a zebra
>>101989007can gen flux tiddies just fine with the right wording and parameters. don't even need a lora. but hey>>101989013nonsense. brotha lower that guidance>>101989017dude I didn't make those rules.>>101989030its a good one. artium also very nice for that kind of stuff.>>101989063try forcing the clip onto your cpu/ram. loras need tons of additional vram
>>101989063probably a bug with the lora loaders or loras on quants. try the fp16 or fp8 model see if it happens then.
After updating comfyui now it loads 60GB into memory
>>101989067Porn loras always suck and are a pain to use. It'll take a good nsfw finetune to properly uncuck Flux
>>101989079>loras need tons of additional vramthat shouldn't be the case
>>101989033>Implying 28 is enough
>>101989062>SD3 2B is a badly/undertrained trained piece of shit, do you think they WANTED for it to be that bad?no one care what they wanted retarded faggot, they nerfed their model at the end of the day, and they knew that removing any kind of nudities in their dataset training would lead to a nerf in the capabilities of their model, they already eaten that experience with SD2.0, so yes dipshit, they wanted it to be this bad
>>101989062>Companies avoiding being associated with pornographic content is an old taleand they probably also don't want to be associated with "dangerous" hyper realistic content. they don't want government regulation man, it's not that complicated. they want their ai gens to be obviously ai.
>>101989004>>101989049damn I didn't see the spatulas to! the guy has a serious fetish that needs to be unlocked
>>101989062kill yourself
>>101989067>its not curious at all, nipples are censored in the datasetWRONG WRONG WRONG WRONGwe see DIFFERENT results on tits.That makes no sense. It "blobs" tits. It's not a consistent censor, this means that the censorship is something strange. It's not an algorithmic censorship on the training data.
>>101989062it's gotta be debo right? only him can be this retarded, too bad he hasn't used one of his faggoted pictures so I can't filter him that way, can you please put your pictures again debo? I don't want to hear your retardation at any time of the day
>>101989111brainlet cope, thats not how a diffusion model works, it has no internal algorithm that censors things, its in the weights
>>101989098>>101989110>>101989118actual retards with retard arguments, go back to school>>101989099and you think freely offering models that gets you 95% of the way to realism (when you only need 50% to fool the average retard on Facebook) is avoiding that controversy?you're all idiots
>>101989063apparently there are some issues with loras: https://github.com/comfyanonymous/ComfyUI/issues/4366
>>101989092but thats how it is. if I dont force the clip into the ram I max out 24gb vram and everything goes to shit. with one lora. technik die begeistert.>>101989105NOT SURE IF SCARY OR CUTE
Question. Does using a LoRA make gens take twice as long for everyone else or have I fucked up my workflow somewhere?
>>101988925Sir that's illegal
>>101989140something's off, it is either sending users into lowvram or causing oom crashes otherwise. >>101989135
>>101989135I hate the way lora is handled on ComfyUi, it unloads the model everytime you wanna change something, it completely ruins the fun, it was way smoother adding loras on A1111
>>101989157No President Trump made it legal, clearly
>>101989140no, you raised your cfg? that would double your gen time. no need to, leave it at 1 and adjust the available parameters to taste.
>>101989135>>101989162Yeah it's doing that for me on my 4080.The LoRA I trained is only 18mb too, not some retarded 1GB LoRA I downloaded.
>>101989111Models are pretty good at generalizing. You don't need anything more than male topless pictures in the dataset for it to be able to generalize nipples imperfectly to females.
>>101989186It doesn't :^)Get it? It doesn't, though it should. Male tits look fine in Flux.
>>101989127>avoiding that controversy?yes? it's defienetly going to be way less compared to releasing a model capable of non slopped realism. why do you think every ai company is so fixated on "safety"? ai is new and they are trying everything to stay ahead of the competition while playing it safe to avoid government regulation. like i said before, it's not that complicated.
>responding to retards or llmsimagine
>>101989195>compared to releasing a model capable of non slopped realismYou still haven't shown that they CAN train such a model and until then your hypothesis is baseless.
>>101989166No cfg is at whatever it is with this node, there's nowhere to set cfg with it, but even with the basic ksampler it's the same, and goes to normal speed with no LoRA attached and nothing elde about the workflow changed.
>>101989214just wait a while for the finetunes and then we'll see if i'm right or wrong
Schnell is alright
thread so cozy when anons argue
>>101989222You think finetuners are gonna get hit by the cOnTrOvErSy?
>>101989246how am I supposed to fuck it?
just gonna gen topless ella purnell until I die>>101989207funky style, what was the prompt?>>101989220there should be a "basic scheduler" attached to the sigmas input of the sampler, there would be the CFG value.>>101989228good face. Need to back to it.. bit too large a file to just take a nap on my HD forever
>>101989243won't matter because they aren't a large company that depends on their models for profit
>>101989256Get a power drill
>>101989257>there should be a "basic scheduler" attached to the sigmas input of the sampler, there would be the CFG value.It ain't. It's not the CFG.
>>101989256By using the drill dick premium 9000, only for an extra 9999$
>>101989259then how does it prove you right or wrong?
>>101989256Buy an onahole upgrade
>>101989235At this point there is always 2 anons arguing every day here about a random useless topic. So yeah I guess it's comfy because it feels familiar.
>>101989136>NOT SURE IF SCARY OR CUTEits antartical flumpuszuch
>>101989246>>101988894kek and very goodcan you make the head round and transparent and body skin color
>>101989268right, there isn't. sorry I am sleepy. hmm. try a different lora? I've really just started testing flux with loras but so far I see no gen time increase just by adding one.>>101989294lol. they probably look really good once upscaled.
>>101989273would. should be modular and universal onahole format
>>101989257huggingface so lost the prompt, it was something along the lines of "poor crayon drawing of a dumb-looking man drooling with a speech bubble"
>>101989272if finetunes are capable of non slopped realism then it means the technology is more than capable of it too. the reason why the base model looks so sloppy is because black forest labs intentionally nerfed their models to make it's gens obviously ai. that is my hypothesis. unlike finetuners ai companies have alot to lose from government regulation and controversy.
>>101989174does the same thing happen when you use the fp8 model? or just the quants?
>>101989319Happens with every LoRA. Wondering if I have fucked up my workflow in some way. I moved some outputs around.
>>101989332the realism might improve but only on those things it is finetuned for, the model will be less general
>>101989334Yeah I tested that yesterday, same thing
>>101989367So I guess the problem is based on the ComfyUi software and not so much on the loaders
Massive fucking titties!
>>101989371it could still be on the loaders, just on comfy's side. honestly at this point it's better if we get a node to convert loras to our desired quant, then we just load the quant lora with the right model.
celeb loras are all over the place, meh.>>101989344what I saw on your images looked ok. not much to fuck up, really.I am using rgthrees power lora loader btw, just for reference. and the dev model, fp8.
>>101989403Here's the workflow anyway, maybe you can see something I don't. https://files.catbox.moe/2pub6t.json
I'm surprised Flux could do Sailor Moon that well, that anime isn't that well known and it's old
>>101989428>that anime isn't that well knownWhat are you, 20? This shit was MASSIVELY popular.
>>101989438>wasyeah, was, but nowdays who talks about Sailor Moon seriously?
>>101989428>that anime isn't that well known and it's oldI hate zoomers
>>101989428>that anime isn't that well known and it's olderm...
>>101989319>lol. they probably look really good once upscaled.yep, good upscale material, pic related
>>101989447Does it matter who talks about it? The training set is massive. I guarantee it know a lot of actually obscure Anime, it just likely isn't tagged with the anime name. That's why it trains so well, it already knows what it's looking at, it just needs to be told what it is.
>>101989447sailor moon is kind of like mario and sonic where people who don't know what it is recognize the characters. obviously not at the same level but still, very popular.
>>101989465>it already knows what it's looking at, it just needs to be told what it is.meaning embeddings could be very powerful
>>101989481for some reason the embedding side of things is being ignored atm, maybe when things cool down a bit.
>>101989465>Does it matter who talks about it? The training set is massive. I guarantee it know a lot of actually obscure Anime, it just likely isn't tagged with the anime name.that's weird though, Flux doesn't know a lot of shit, it can't fo Franklin from GTA5, one of the most popular series of all time but it can do Sailor Moon?
>>101989488>>101989481It's a shame nobody is trying that, LoRAs are king because they are well used already. It's hard to make people change the familiar.
>>101989425testing it now. dont have that lora but I am also seeing 2.4s/it, thats not good, should be 1.0x.one moment.
>>101989428>that anime isn't that well known and it's oldthis is insane bait
>>101989497how many gta5 screenshots tagged with specific characters do you think are in the dataset?
>>101989425Have you tried connecting the model output of the lora node to the ModelSamplingFlux node?
>>101989497I guarantee it does just not his name, because the VLM that tagged the images of Franklin did it like this
>>101989512that's the thing, if they haven't decided to purposely censor their dataset, I find it hard to believe they found more screenshot of Sailor Moon than GTA 5
>>101989497it knows what the llm being used to tag it knows. it's not like dalle3 where they hired thousands of pajeets to manually tag the dataset on top of the llm tagging.
>>101989523it may be hard for you to believe, however, there are certainly more images of sailor moon than well tagged screenshots of gta5
>>101989522vs thisThis is the only reason it doesn't know these concepts.
>>101989532>to manually tag the dataset on top of the llm tagging.other way around, they made a very well tagged data set to train their caption model to then tag their huge data set.
>>101989428I just punched my keyboard because of this, had to pick up all the broken off keys from the floor. Please watch what you say in the future.
>>101989523the vlm knows sailor moon >>101989536doesn't recognize franklin
>>101989536that's fucked up, didn't know the VLM would precisely talk about those concepts like Sailor moon and then decide to ignore to say it for gta 5
>>101989542you need to take some pills anon
>>101989547it depends on the VLM of course, we don't know what BFL used
>>101989508
>>101989547the vllm hasn't made a conscious choice to ignore details for gta5, it simply does not know those details
>>101989564like how is that possible? GTA is everything but an obscure concept, even my grandma knows what gta is
>>101989425ok i exchanged the gguf model loader with the other one and tried again with the dev model (fp8e4 bla) and voila, full speed. and here, just for reference, the wiring. its GGUF related
>>101989547It's full of weird things like that. I doubt they used Joy Caption specifically but the same basic principle applies. It 100% was trained on Britney Spears. And yet...
>>101989536Yeah so anyone who is finetuning flux in the future please add the names to any auto captions you make, it doesn't take long, you can select all images of the person in Taggui and simply add the name to all of them.
>>101989428Successful bait Everyone knows sailor moon
>>101989574sailor moon was an iconic character, gta is popular but franklin isn't as iconic. sorry franklin
>>101989574The underlying LLM might know about GTA in detail but the VLM might get confused and describe it as a photo without making the association with GTA as seen here >>101989522
>>101989579Also it would be an interesting experiment to add random names to any random people you add to the training, maybe it will help avoid same face syndrome because you are training all kinds of names to different faces, so you can change a lot of looks by using random names when you gen images.
>>101989575Testing
>>101989574because you havent labelled gta 5 screenshots with the character names
>>101989579>tagguiqrd?
>>101989597yes, please, so many loras are just "woman" this "man" that and are impossible to use with regional prompter in attention mode because they leak
>>101989603https://github.com/jhc13/tagguiI really like the design, makes tagging super easy. Although I hope they add joycaption to it one day
>>101989603There's a box that you can add a tag to every image on training guis. You just write the tag in and even if an image isn't tagged the trainer tags it with whatever is in the box
Why did faggots raise cfg in the negative prompt version, again? it's tripled the time I have to wait.
>>101989618CFG doubles the gen time regardless of what value it is, except for 1.0 which disables it.
>>101989618To get negative prompting working on Flux.
>>101989612>>101989614ty i am still a vramlet that can't train anything at the moment. the hardest thing to do is find any documentation on how to do these things.
Flux can't make a gatling gun?
>>101989635It's a modern world problem. All the documentation is in video essay form and I fucking hate it.
Is there even a single decent looking flux style lora yet?
>>101989648Use AI to turn videos into articles.
>>101989618to get better prompt understanding, that's why CFG was invented in the first place
>>101989657what is cfg anyway?
>>101989655Only works if the video is captioned.
>>101989649I saw some
>>101989665Cum filled girl
>>101989667Whisper
>>101989648>All the documentation is in video essay form and I fucking hate it.desu I prefer when it's in a video format, feels like having someone near me explaining how to get shit done, text only is so cold and doesn't have the nuances a video can have
>Hatsune Miku playing chess with Sailor Moonkek, concept blending is still a thing in Flux I guess
>>101989669nice one anon, prompt?
>>101989674I just use chatgpt now, and this is how we will be getting this kind of information in the future.Can feed it a bunch of articles and white papers and turn it into useable information.
>>101989575This made it significantly worse.
>>101989536>Sailor Moon from Sailor MoonIt's Usagi Tsukino you stupid model!
>>101989698Wait I lied. Still not 1.10 s/it which I get without but a full second better. Output is much worse though compared to Q_8
>>101989704it's only 8B, please understandimagine a llama 3.1 405B VLM
>>101989704>It's Usagi Tsukinobut the form she has on the picture is the Sailor Moon one, so the model is right
>>101989721forgot pic
>>101989704>Usagi is Sailor Moonschizo moment
>>101989721>Output is much worse though compared to Q_8yeah, Q8 is really close to fp16 in quality, you can't go closer than that unless you go for Q8_K but it's not being implemented by llama.cpp yet
>>101989704that's her real name
>>101989723No she isn't. She is in her schoolgirl Usagi Tsukino/Serena form.
>>101989497
>>101989689hyper realistic professional photo of Sailor Moon in sunglasses and cigar in her mouth holding huge stacks of cash, fisheye effect, gold bars and stacks of dollars in the background
>>101989737Well let's see if it whores up my VRAM and needs me to reset it every few gens like Q8 does. For my current purposes it's fine.
>>101989744oh yeah you're right, my b
SELLING TOPLESS VALORIE CURRY PICS BROS. SHE LOOKS... HOT. much better than the real thing. 10 bucks/catbox. anyone?>>101989698thats... uh.. you spilled over into ram?>>101989721hm. but why did I get those crappy 2.4s/it with the GGUF quant then? oh man>>101989745lol
>>101989649yeahhttps://civitai.com/models/647940/flux-atilessence-lora-test?modelVersionId=724910https://civitai.com/models/659029/sxz-elden-ring-flux-lora?modelVersionId=737407https://civitai.com/models/649031?modelVersionId=726131
>>101989749where is q8?what gpu do you have?
>>101989758
>>101989760>sxzlast time I tried his 1.5 things it was the friest frying fried shite I ever used
>>101989758>hm. but why did I get those crappy 2.4s/it with the GGUF quant then? oh manI got that samish speed using a LoRA on Q8. Using it on FP8 is giving a steady 1.74s/it. >thats... uh.. you spilled over into ram?Just pre-gen shit. Seems to happen a lot when first loading a model
>>101989748Thanks anon
>>1019897614080. https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main
>>101988878Doubtful.
GGUF doesn't increase my iteration speed so I'm back to my favorite model.
>>101988842This is simply untrue. I have trained a LoRA of myself and genned images of it. People who have known me my whole life, and people who look at me every day, cannot tell that it was AI generated. I even changed my LinkedIn profile pic to an AI generated one because I'm not buying a fucking suit.
>>101989796More like IQ15 text encoder
>>101989804she spilled a lot of strawberry jam with her unwieldy saw hand
>>101989789Thanks.Much better ai card than my 6950xt.
>>101989814same
>>101989796What's your iteration speed?On dev I get 13 seconds per iteration, if I use the basic setup, cfg 1, euler.
>>101989814>I even changed my LinkedIn profile pic to an AI generated one because I'm not buying a fucking suit.dare I say based?
>>101989814>because I'm not buying a fucking suit.You can buy a suit at a used shop, it's no excuse.
>>101989844I'm not wearing clothes someone else has masturbated in.
>>101989760>https://civitai.com/models/647940/flux-atilessence-lora-test?modelVersionId=724910I really love that lora, brings life to the generic styles Flux is having in
>>101989833For 1024x1024 53s/it on Comfy and 47s/it on Forge.
I don't know.
>>101989783
>>101989892>53s/it on Comfy and 47s/it on Forge.
>>101989878Those colors hurt my eyes
>>101989892My condolences, anon, I get 2.4s/it and it feels too slow for iterative work.
>>101989771same thing, but it's hard to fuck up a style lora on flux, no more cope settings. just get a diverse dataset and you are good to go
>>101989892mine aren't that slow but even that was enough to get me to cave for a 4090 which is awaiting its case and psu to arrive
>>101989939Isn't she underage?
Fuck it crashed again, there's definitely something wrong with how ComfyUi handles loras
>>101989814yup, this is kind of my point. if some guy can create photorealistic gens with a simple lora why doesn't flux base with their millions in funding? the technology is more than capable of it, it's that the companies behind these models intentionally avoid it to prevent regulation and controversy.
>>101989771Yeah like that other anon said, it's harder to mess up with flux, loras work so damn well.
>>101989721>>101989776Yeah this definitely fixed the issue and I realise the speed is because I am genning at 1024x1536 instead of 1024x1024 so really it's correct. It's also maxing out my RAM and making me unload every few gens.I really really like Q8 but it's tedious to work with. I hope we get some improvements soon.And yes, Q6_K did the same shit with needing to unload the model.
didn't know there was a blonde miku
>>101989903Can confirm, that's roughly what my 1080Ti does for that resolution, that's why I've been using 576x576 and it's still a 3 minute long pain.I really need to grab a used 3090 to play with this, I'm demoralised with these times
>>101989895
>>101989954It's unlikely intentional, Anon. It's not some "let's make everything look AI" thing. When you're training on billions of images, it's going to look a bit sloppa as the things all converge. When you fine tune it understands a single concept really well.
I'm glad Miku finally found a friend to play with
I want to see people very first Flux gens.
>>101989995
>>101989995that was my very first one, I tested out Comfy's workflow kek
>>101990005What happened?>>101990014I guess I should specify non-comfy gens. 150 fox girls is unnecessary
>>101989995I want to see your next few Flux gens after that.
My teacher at school
>>101989995would be this
>>101989981sorry anon but i'm going to standby what i think, openai does it with dalle and i don't see why black forest labs wouldn't do the same. image gen is a finicky topic and i think it would just make sense to play it safe. it's my opinion and i respect yours too. let's not continue this.
>>101990026From a different prompt or the same? Here's my first different prompt.
>>101990026>>101990054And the next one after that
>>101990062Yes miss, I am a bad baaaad racist. You should punish me.
>>101990023I used karras with 20 steps.
>>101990066And after that it's just a shit load of nazi catgirls>>101990083Yeah that'll do it
>>101990053you're retarded, you lack critical thinking
>>101990091he's right though, image models are a really controvertial thing, it's not a coinscidence the US government is talking about adding some regulations to it, it's a serious thing
>>101990091next time you take a dump i think you should eat it instead of flushing because you're probably shitting out your remaining brain cells everyday
>>101990106no, he isn't right about why models aren't yet the super duper perfect reality engines he thinks is already possible>>101990111okay retard
>>101990121>why models aren't yet the super duper perfect reality engines he thinks is already possibleyou know the answer, if they make it too good the government will make it illegalhttps://www.responsible.ai/a-look-at-global-deepfake-regulation-approaches/
>>101989995It was le galaxy in a bottle
>>101990141first you have to show they CANmake it "too good", then you can reason about why they don'tthat's reasoning fucking works you fucking illiterate Appalachian redneck mouth breathing cocksuckers
>>101990165>first you have to show they CANmake it "too good" >>101988745
>>101990165excuse the typos, my jimmies are rustled>>101990173you call that "too good" you fucking dipshit literally kill yourself RIGHT NOWYOU'RE STUPID, KILL YOURSELF
Lots of discussion as usual and so here's the next fresh bread...>>101990155>>101990155>>101990155
>>101990181keep shitting your panties and you'll loose the few precious brain cells you have remaining up there. calm down.
>>101990220you're still retarded no matter what I do or say
>>101990054the same i need to cum to cleopatra
>>101990236say that again, i dare you.
>>101990262you're still retarded no matter what I do or say
>>101990289thank you for following my orders like the good little doggy you are. bark for me next.
>>101990252It's Nefertiti
>>101990329I can't cum to this but thanks anyway. The comic stuff looks good, maybe I'll gen some of that myself, flux's anime still feels a bit mediocre
>>101990325it was a dare tho
>>101990053you were never looking for discussion anyway, you came to the thread with an opinion and just repeated yourself like 12 times. fuck off already.
>>101990409>you came to the thread with an opinion and just repeated yourself like 12 times. fuck off already.and you didn't do that aswell?
>>101990409>just repeated yourself like 12 timesi was repeating my point. i was repeating what i genuinely think.
>>101990387i dare you to shit yourself right NOW!
>101990445I now know your dares aren't genuine, therefore you a man without honor and retarded. Don't talk to me again
>101990470sorry, i my dare back
take*
annoying, tried Forge to compare with Comfy, same prompt and seed gives different results on both
>>101990496try smZnodes comfyui extension which has a "Clip Text Encode++" node, allows you to choose a1111 parser and turn normalization on/off (this is the token weighting calc system i believe) and multi conditioning (for use of AND instead of breaking stuff up into silly conditioning concat blocks). AFAIK forge should be the a1111 parser, don't know about normalization so try both ways. but also 1:1 recreating other people's outputs or your own old outputs is not that much of a priority so don't worry too much about being able to "test" stuff by replicating others' shit; half the time people upload stuff with metadata that's incomplete relative to what's needed in terms of loras or multiple upscaling steps anyway.
>>101989995>>101990023
>>101990496forge is way faster
>>101988457sex
>>101991134techlet cope
>>101989852Nobody masturbates in a suit, too hot.
