Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101833248>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
TY baker
Niggerlicious general
https://civitai.com/models/638665/shimoe-koharu-flux-dev-lora?modelVersionId=714197ITS STARTING BROS
>>101836492>>AuraFlow>https://fal.ai/models/fal-ai/aura-flow>https://huggingface.co/fal/AuraFlowsisn't it about time this gets removed from OP?
>>101836534FUCK YES
>>101836477It is garbage, it's not quantized, it's a naive fp16->fp8->fp16 conversion, precision is lost
Cursed thread of samefaggotry and shitposting.
>>101836536What happened?
>>101836248Inpainting also doesn't seem to work properly, no matter the settings. All it does is blurs the area and adds a weird halo
>>101836489Rent free
>>101836588looks pretty beta to me. first test run I set it to "gen forever", cancelled and it just kept going without progress bar, had to kill it. well glad I left reforge as a backup.
Blessed thread of frenship
>nf4 fp8 fp16I have a 3090. do I go with fp16 for flux or lower?also, is it possible to train a lora for flux with onetrainer?
>>101836648Well it's pretty good since the pedo is not here spamming it with bikinitoddlers.
>>101836443flux dev without changing the weight_dtype
>>101836712
>>101836492>>101835305>>101836461ty baker
got NF4 working in comfy, now to see if it works fine with loras.
>>101836662Pregnancy is the most natual thing in the world. If you think about cumming looking at this image then....
checked flux and still believe 3D diffusion is the future for both of video creation and single 2D image creation.I am not talking about SAI or something specific.
>>101836534We're in the process of becoming so so very back!
Seeds 0-3https://imgsli.com/Mjg2NTQzhttps://imgsli.com/Mjg2NTQ1https://imgsli.com/Mjg2NTQ2https://imgsli.com/Mjg2NTQ3Next up is nf4 vs fp16. nf4 > fp8 for sure, or at least fp8_e4m3fn>>101836712based
Real talk, CatJak is a twitter tranny faggot that ruined /sdg/.He's a massive drama queen and blogged on /sdg/ about his coombrain lora training and got BTFO by the thread and got extremely butthurt, so he now has a vendetta against that general and swore to destroy it.Now he just shit that general up, samefag and try anything in his power to make everyone leave that place.This is why /sdg/ gets flooded with same posting about PW, debo, avatarfags and "schizo anon"This is why he deliberately starts flamewars and /sdg/ and redirects anons to his fake general, while trying his hardest to create "friendly" and "schizo-free" atmosphere at /ldg/ .He hates /sdg/ but can't get rid of it seeing as they're 99% of anons see through his bullshitHe's also called "ranfag" btw
>>101836860deja vu
so, was the controlnet for flux a nothingburger?and is it working with schnell/comfyui?
>>101836843kek, fp8 chads we lost
>>101836843pumpkin in nf4 instead of apple banana
>>101837011
nf4 vs fp16https://imgsli.com/Mjg2NTQ4https://imgsli.com/Mjg2NTQ5https://imgsli.com/Mjg2NTUwhttps://imgsli.com/Mjg2NTUyThese are all with euler, 30 steps, simple scheduler, 3.5 guidance and fp16 text encoder btw.I'd say fp8 <<< nf4 < fp16>>101836937It's over.
>>101837061Cant get any of the existing flux loras to work with nf4, have you had any luck?
>>101836843>>101837061MiusexMigusexMilusexMipusexMibusex
>>101837061fingers and hands tend to look smaller with nf4, but overall I like the faces more in ng4
>>101837061>>101836991>>101836971>>101836937>>101836877Put your filename back debo.
nf4 animagineXL when!!
>>101837101kek>>101837088No luck with loras, guess we just have to wait until someone makes them work with nf4
>>101837061OMG IT MILU!!!
>>101837164Just had to make sure it wasnt just me kek
I'm a little confused: even if NF4 was worse than FP8, it still might be worth running for VRAMlets, so it's really impossible for it to "lose" unless it was slower or took more memory than FP8 or if the quality loss was intolerable. I mean, quantization is probably going to hit some tasks more than others. If there was no trade-off, it would be a "free lunch," and those are hard to achieve. I still think it's worth doing side by side comparisons.
>>101837229>I still think it's worth doing side by side comparisons.Anon...
>>101837229Nothing strange has happened.It is quite similar to the history of compression algorithms or video codecs.
https://civitai.com/models/638904/anthroid-flux>hyperswole fetish lora came out before the big breast lorakek
>>101837385it's 69
Wide
>>101837061Does it always generate all the elements for you? This is with T5 in float8 and unet nf4
Tall
>>101837414>>101837444detailing on these are so nice considering they're not upscaled/resampled
need flux cocks asap
>>101837430I never had an image missing the bottom caption/changing the language/put the sign text on her shirt, so I would guess it's because of the fp8 T5
>>101837397Correct!
>>101837490neat
I am trying some of the new loras that have come out and I am confused. They do things, but none behave as expected. I am using guidance 3 Lora strength 1 and flux dev and fp16. Am I doing anything wrong?
>>101837488Yeah, that was my guess too. Thanks.
>>101837490use the realism lora to remove the plastic skin look
https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4that exclusivity didn't last long I guess
>>101836492Since flux relies on long prompts and has good prompt comprehension, why aren't people training text embeddings for it rather than LoRAs.
>>101837390The lora still has weird nipples... kek
>>101837567the popularity of embeddings died with XL
>>101837564actually better because you can use fp16 t5 with it too
prompt: a big nigga
https://civitai.com/models/638006/yukihana-lamy-lora-for-flux?modelVersionId=713889Nipples still look kinda fucky
Hmm, trying Flux for the first time (nf4). Generating takes ages since vramlet but it works. However the quality seems kinda ass. Should I just increase step count or what's the deal with the artifacting I see here.
I'm very impressed by this.No model so far has come even remotely close to being able to generate someone using a bow, without the bow becoming an abstract piece of art.
>>101837642Not just the nipples!
>>101837642what we really need is a good hentai finetune but who knows how long til that happens
>>101837683nice gen, almost perfect hands
Anyone else get an OOM while using a lora with fp8? Oddly it seemed to work for 1 gen, and then it didn't. Also I restarted Comfy and tried again, and it seems like no matter what, I can't get it to not OOM anymore even for a single gen. I'm on a 3090.
>>101837722Yeah this model kicks ass, it's leagues above anything else as a base model.Sure for the moment it's a bit stiff and doesn't want to make certain things, but now we just wait for some autismos to start finetuning this and we're getting pretty damn close to the commercial models out there.
>>101837774I don't think it will be worth even trying to use loras for another couple weeks, shit will probably be optimized by than. Seems like most of them dont even work anyway.
I'm a nodelet and only swapped for Comfy for Flux, is it okay to go back to Forge now?
>>101837835>I'm a nodelet and only swapped for Comfy for Flux, is it okay to go back to Forge now?
Can flux write words it wasn't trained with?Random letters?
>>101837869
>>101837532~Thanks I will do that when they we figure out how to use loras with nf4
>>101837869I can't make it write words in foreign scripts, it just ignores the characters in the prompt.
>>101837835I've got Comfy and Forge both installed and use both (I like Forge better for inpainting) and with the Comfy NF4 custom node linked above, Comfy is now faster at running the NF4 checkpoint than Forge by about 4 per gen at 1 megapixel on my 3090.If you're okay with waiting those extra seconds in return for not having to deal with nodes then Forge will work fine for you though.
Wtf now it works again. What the hell is going on.>>101837828Yeah, guess there's some bugs.
>>101837922*4 seconds, that is.
since it's so hard to finetune, can people basically keep merging loras into Flux to create the ultimate model? Or will that ruin the quality?
>>101838047>Or will that ruin the quality?One will start to overwrite the other since they are being trained independently of each other
So is miku the new Emma watson
Reminder that DALL-E 3 has all your favorite characters and celebrities in it already.
NOO NOOO RUN AWAY NOOOOO
https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4/issues/4so the vramlets that were supposed to benefit from this are fucked if they use the comfy node instead of forge. guess i will just wait until forge supports the fp16 t5
>>101837922>>101837937Perfect, thank you. The minor extra gen time doesn't bother me
>>101838119Good luck getting them through the filter. I can always use SDXL if I want boring celebrity 1girls
Are ControlNet IP Adapter models borked on ReForge or is it just me?
You guys can we get some hot weather reporters here please?
>>101838157They wouldn't be boring with DALL-E 3.
>>101838135Would make a good RE2 Remake mod>>101838171Yeah sure thing
>>101838185They also wouldn't look much at all like the celebrity you're trying to gen
>>101838135>its the evil rape man (Rape mans twin brother)
>>101838119they don't have Rape Man tbqhwy
>>101838047doesn't work like this, they will overwrite the knowledge, not add to it
>>101838210doesn't finetuning kind of do that when you are fintuning an already trained model? I have no idea though
>>101838209kek'd
>>101838201Imagine choosing to only remember DALL-E 3 as the locked version it currently is and not the greatest image generator it was and will once again be when it is freed from OpenAI's clutches.
>>101838233What do you mean by freed?
>>101838223no, different processyou can bake one large dim lora and merge it into the model, that one can work, but not many small loras
>>101838233>will once again be when it is freed from OpenAI's clutches.by the time that happens local will be better than it
>>101838233Likenesses were still piss poor compared to a SDXL lora specifically trained on a single person
Can flux handle violence? Like punching, kicking, choking etc
>>101838262OK thanks.
>>101838171Here you go
>>101838233I remember when dalle was released and it was amazing, but flux is almost that level except for styles. it's very fun for memes/shitposting
>>101838281Now that's a hot beard.
>flux booba lorahttps://civitai.com/models/639046/flux-booba
>>101838191>>101838298good gravy please gen some nude ones and catbox them, this is hot as fUCK
>>101837662I like the retro anime-esque throw back.
>>101838256leaked>>101838267DALL-E 3 already has great likenesses of many characters and celebs, were you not around when it was made public?haven't seen a single good celeb XL lora, the likenesses always fall apart with more complex prompts or when you want to use them with NSFW finetunes>>101838288I tried making a worm's eye view shot that DALL-E 3 was so good at with Flux and it just doesn't understand it, no matter how I phrased it it just can't make extreme low angles.
>>101838358DallE will never be leaked
>>101838358Theoretical uncucked Dalle is better than flux but honestly I wouldn't be too sad about it. We're extremely lucky to get a model like this already. I was pretty sure local would stagnate for much longer than this.
>>101838409Bet.
>>101836843Was she supposed to be a real person or not?
>>101838451Yes>A photo of a Korean woman holding up a sign that says "FORGE OR COMFY?" Her right arm has bracelets on it. Her eyes and hair are black, her t-shirt white, and her pleated skirt blue. There is a subtitle caption at the bottom of the image which also says "TO NF4 OR NOT TO NF4?" The girl stands leaning on a brick wall outdoors. The weather is slightly overcast, giving the image a blue tint. The skirt is slightly torn up. On the girl's t-shirt, there is a moe drawing of Hatsune Miku's face. The t-shirt says "MIKUSEX". In the background, there are bowls of fruit containing apples and bananas. Some of the apples have eyes and are looking at the camera.
>>101837525This seems to work for me. >>101836534
somewhat surprisingly it doesn't know the classic playboy bunny outfit
>>101838623suprisingly it knows actually a lot about womans lingerie.https://github.com/ataylorm/FluxAIGridComparisons/blob/main/LingeriePrompts/FullGrid.jpg
>>101838268
>>101838333What I got so far. Feels like it loses a bit of quality when trying to go nude, but not bad.https://files.catbox.moe/d3nrlj.jpghttps://files.catbox.moe/d5vsvr.pngWarning this one has white paint on her bodyhttps://files.catbox.moe/n74jlf.png
>>101838668Flux can't do Gumbo Slice, this squarely puts it beneath Midjourney and DALL-E 3 in the ranking.
>>101838268>>101838746
>>101838753Gumbo is larger
Can someone translate into retard for me. How can nf4 be significantly smaller yet produce comparatively similar quality results as the full fp16?is nf4 just much better at calculating shit? How could this be possible with so many less numbers?
>>101838785Separate sets of numbers allow it to scale back up without losing precision
>>101838785>produce comparatively similar quality results as the full fp16Is that certain yet? I haven't been keeping up with the comparisons discussion.
>>101838785calm down only 1 or 2 prompts have been tested lol
>>101838785some weights are very important and some are not, keep the important ones at full fp16 (or even fp32) quality and the unimportant ones at much lower precision and you end up with something much smaller but not much worse
>>101838785fp16 = uncompressed filefp8 = zip archivenf4 = 7z archive
>>101838813>Is that certain yet?From my own tests, yes, especially if you use the fp16 clip. >>101838824This analogy makes the most sense to me, thank you.
>>101838636yeah, I've noticed
nf4 T5 when?>>101838838stop calling T5 "clip"
>>101838753Not really
>>101838855It's very hard, Anon. After so many years of using a specific terminology shit just becomes normal, whether it's correct to say it that way or not. As long as you understand my intent that's all that matters really.
>>101838824you're going to get absolutely ripped apart if this analogy isnt accurate(not by me i dont know anything)
A screencap I took of the news back in the day
>>101838882>As long as you understand my intent that's all that matters really.that's literally how you get the current usage of literally
RAPER
>>101838745muh dik these are pretty good even if censoredboy those porn models are going to be INSANE.>>101838891fffuuuuuuccccckk tempted to take a look at this image on a crt to see how "authentic" it gets.
>>101838899why can't she just leave Rapeman the fuck alone, what's wrong with liberals
>>101838668>>101838753I thought hands and feet were a solved problem already. What happened anons?
>>101838908this place is just full of shit prompters and retards
>>101838901Yeah when this gets some good loras and finetunes...oh my. Gotta be patient.
>>101838908>What happened anons?new good free model released
>>101838928Care to demonstrate how it is done?
>>101838907YEah the reality is Rape man is just trying to walk home from work (where he rapes people all day) and these damn liberals keep stopping him and take selfies to try and ruin his image.
>>101838916put filename back debo
>>101838908>thought hands and feet were a solved problem alreadythey are with DALL-E 3 and somewhat with PonyXL
>>101838952I'm not debo bro
>>101838896If you're going to be pedantic about this shit at least use proper punctuation.
shitty prompter containment general lol, truly cursed thread
>>101838967no go fuck yourself
>>101838966>debo-tier gen>writes like debookay thread schizo who the fuck are you kidding
>>101838973>Whining and not posting any imagesDon't bother, I already accepted your concession
>>101838973Very sorry that our gens are different to the last ones we posted, and different to the ones from the thread before.We appreciate your feedback and will endeavour to post the same shit over and over again, learning nothing.
>>101838989Ok I'm not going to argue with you I'm just posting gens man
can't wait until we get fp16 t5, fp8 t5 fucks with guns too much
>>101839015Okay debo
>>101838137>vramletsfixed now apparently
>>101839022why can't you use fp16 T5 right now?
>>101839036Can't do it in Forge.
>>101839036He's poor.
>>101839022yeah from everything I've seen, downgrading to fp8 t5 is far worse for quality than going from fp16 to nf4
>>101839043you can use nf4 in ComfyUI
>>101838973blessed thread of frenship
>>101838973>has problem with shitty thread>stays in shitty threadthe devolving psyche of some posters is more impressive than generational leaps in AI
>>101839043just use comfy? why use training wheels ui
>>101839054Sure can, but he is clearly using forge. I am having great results with it in comfy but not everyone wants to use it, and that's fine.
>>101839071he should learn
>time taken per gen on a potato 4GB gpu with 16GB RAM in comfy at fp8, 20-25 minutes>time taken per gen on potato 4GB gpu with 16GB of RAM in forge at fp8, 6-8 minutes>card doesn't support NF4 unfortunately>model loads faster on forge, computer is still usable while it runs in the backgroundalso pro tip for forge, if you want to use schnell or schnell merges that refuse to load, just go to backends/huggingface/black-forest-labs, make a copy of the FLUX.1-dev folder and name the copy FLUX.1-schnell and boom, Forge can load use Schnell.
>>101839070simple people want to gen simple woman with simple toys r us UI
>>101838358It's possible getting low angles out of it depending on the subject in question and even then you might have to be somewhat creative coaxing it out.Something like trees viewed from below is such a common subject Flux is able to do it just fine.
>>101839071>he is clearly using forgeWhat would be the reason for such retardation? Everybody is using ComfyUI.
>>101839022>can't wait until we get fp16 t5, fp8 t5 fucks with guns too muchcan't nf4 do guns well?
>>101837900lol>>101837915OK I guess the limitation is random combination of letters and foreign scripts then.
>>101839090sure but that doesn't help me look under Miku's skirt
>>101839093I am using Windows instead of Linux
>>101839093simmer down comfyanon
>>101838745She looks better clothed.Or in lingerie.
>>101839093go back to /sdg/ cumfart shill
>>101839102You sure you're trying hard enough?
>>101838951where can i apply for this job?
I don't know much about this character but yeah the lora's working well I guess. Fun times ahead brothers.
>>101839093>What would be the reason for such retardation?idk man they are convinced low precision is same quality as high precision
>>101836492flux is bloat
>>101839126Can't see panties, but very nice angle for legs.
>>101839129always like these
Not sure why I bother trying to do fp16 gens I know I can't. I just really want to provide some useful comparisons.If someone with more vram than me and a touch of autism wants to test stuff, I would like to see a few comparisons between fp16 default, fp8 + fp16 t5xxl, fp8 + fp8 t5xxl, nf4 fp16 t5xxl and nf4 fp8 t5xxl
Why are you guys using flux? Emad said SD 3.0 is the last image model you will ever need.
>>101839139this general is so fucking retarded man
>>101839149Thanks
>>101839166both generals are
>>101839153there are some comparisons earlier in the thread, not between text encoders but fp16 vs fp8 vs nf4
>>101839129wanna try something different instead of spamming same prompt for weeks
>>101839184true.grim
>>101839186Sorry you are so upset
>>101839185I know, but I am specifically interested in how fp16 t5 on nf4 compares to fp16 default
>>101839015brutal gen, keep em coming!
>>101839203dude it's not personal, just a suggestion. hope that didn't make you too sad or something but you're just reposting what's essentially a same fucking picture. that's a little bit sad
>>101839126hmm, if I remove "directly from below" or "worm's eye view" it's a normal shot, only seems to work with both
I guess the dataset has a lot of film stills
>>101839261>>101839203
cursed thread
>>101839268Yeah you have to really push it with this, while also describing a picture that makes sense to be viewed from below.It's possible, but requires some luck and imagination.
>>101839300>It's possible, but requires some luck and imagination.DALL-E 3 just does it because why not
Sometimes I forget how clueless normies are when it comes to spotting AI. Go look at the shit being said about the "ai generated crowds" blowing up twitter right now kek
>>101839306because it's not shit
>>101839300Imagine when all of this can be done in a VR simulation of the world with natural human reactions etc.
>>101839306no it doesn't
>>101839113yeah I agree, something about the tease is much better than nude. >>101838901Me in the back.
>>101839306Yeah it's the gold standard, but I'm more than happy with this thing.I seriously thought the AI bubble would pop before we get a new better model and we'd be stuck with XL for god knows how long.And it's just a matter of time until people manage to make this thing more imaginative through training etc..This model clearly understands these concepts, but there's definitely some censorship going on here. It practically refuses to do upskirt shots, but it's still somewhat possible getting them.>>101839335It will happen without a doubt. Then eventually we'll reach a point of replicating reality perfectly.After living tens of thousands of incredible simulations, we'll get bored and ask the machine to make us unaware that we're living in a simulation.Then the cycle repeats, simulation inside simulation. Who knows, we might be in one already.
>>101839391full prompt for this one?
>>101838950yeah sure but I bet you will make fun of me.First you PROMPT THE SUBJECTSo 1girlThen you prompt her body type So MILF, 69 years oldThen you say what she is doing, On knees, BJTHEN you describe trhe backgroundPublic toiletAnd finally you describe the styleso leaked iphone homeade porn footage
>>101839408this would 100% work as expected with an unfiltered DALL-E 3 btw
>>101839406Here you go:a photo from directly below of a woman standing in street in city wearing business suit and very short black pencil skirt viewed from worm's eye view from below (upskirt:1.3) (white briefs:1.3), stepping over viewer, (white lace underwear:1.4) skyscraper cameltoe viewed from above knees crotch
>>101839391Hmm yeah there the only way to keep it entertaining would be to set our own limitations on it, like certain rules to keep things in check.
>>101839149Same.
>>101839162
>>101839431is prompt weighting for T5 even working on any of the interfaces?
>>101839436https://anbernic.com/en-de/products/rg35xxspthese are amazing btw
>>101839421I agree, the first week of DALL-E was fun when everyone was finding all the secret promptsI wish flux had many of these secret prompts. Maybe one day
>>101839320It's annoying the shit out of me because retards are posting blurry pictures of people's hands claiming that they're deformed AI hands, but if you actually look for just a few seconds without trying to psyop yourself, you can see they're just regular hands with the fingers in an awkward position.
flux dev seems to work just fine at CFG 3 without any tricks, imo you don't actually need any tricks like dynamic thresholding. the model just werks at that cfg
>>101839524Setting that aside, there are literally thousands of pictures/videos from that event at every angle and these retards still say its ai. We're so fucked lmao
>>101839574>when the game has no replayability
What's the node for comfy that gives you a gallery at the bottom of your most recent gen? Or is it pre-installed and I just need to enable it?Had to do a fresh install of comfy now everything isn't how I like it
>>101839097it can, i have no issues with guns on nf4 using comfy with the fp16 t5, the fp8 t5 is garbage imo
>>101839607Kek
130 seconds first gen after load 3070 fp16 t5 nf 485 seconds if prompt is changed45 seconds if no prompt is changedthe fp16 takes a while to do its thing with 32gb ram but it is worth it imo.
>>101839445Hard to say. At first I thought that sure the weights seem to work no problems, but now I'm not so sure because the results have some randomness in them.I'll keep on using them regardless until we know for sure, doesn't exactly hurt the gens to throw them in there.
>>101839622
>>101839524yeah that's the main problem that will continue in the future, it will be less about fake images and more about real images being treated as fake. For important matters like war etc...
>>101839765Being an autist who has seen millions of ai images will finally pay off kek
>>101839765It's already happening with lynch mobs going after artists after they accuse them of making AI art, despite not being the case.
>>101839765Also politicians who get caught with images or video of them doing something compromising will escape accountability by falsely claiming it's AI.
>>101839821based belly enjoyer
>>101839840Not sure why these keep coming out blurry though
I just want my celebs back, they have been slowly disappearing since XL, loras are not the same.
>>101839843Try using Heun
Been doing 512x512 images for testing, then when I like something I increase the resolution.1s/it instead of 4s/it The images are actually pretty impressive at 512x512, I was expecting messed up faces and eyes etc but seems passable.
>>101839846Styles have gotten worse (new models all trained on the same synthetic slop), understanding of characters/people has gotten worse, etc...
>>101839821>>101839843Why is she fucking fat?
>>101839015prompt for this one? very aesthetic
>>101839843had similar issues too, just need to dial in sampler/scheduler/step size. try euler &sgm uniform/beta and like 20-26 steps and go from there. should be crisp-o>>101839846yep. on the other hand, celebs can go suck it
>>101839893Because I'm fat, and I believe we souldn't aim higher than our own level. That's why my wife's boyfriend told me.
>>101839871Didn't help and it takes like 2x as long.>>101839893Because I told FLUX that she was.
>>101839884It doesn't have to be this way. Synthetic images/captions can be good quality. Data sets don't have to be purged of known names.The world could be a better place.>>101839916>celebs can go suck itthat's...what I'm trying to do!
>>101839893because anon is trying to be like Pierce Brosnan
>>101839920Please spellcheck before you post.
>>101839717kek
can flux be tested out with webui forge?
With enough compute, you can make anything look like 1.5
>>101839929There is no hope anon, especially when the people who can bake models for nsfw/etc end up being faggots like Astralite who lobotomize their models by nuking styles/characters/etc
>>101839961https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981
>>101839983Can you do indian food next?
So NF4 doesn't speed up gens if you have the VRAM headroom already, right?Same speed as FP16 on a 4090.
>>101839972thanks. am i safe to assume that flux is heavily censored atm?
>>101839129Looks like one of those fake gameboys with shampoo in them
>>101839374Catbox?
>>101839995out of curiosity what kind of speed does a 4090 give you?
>>101839995you just wanted to brag
>>101839995I go from 1.19s/it on fp8 to to 1.09s/it nf4 on a 4080. I'd imagine the benefit will be negligible on a 4090.
>>101839995No NF4 doesn't speed anything up on my 3090 either. But I'm still using it in Comfy because it lets me use alll the saved vram to load in SDXL models and controlnets etc. for upscaling and refinement. Shit's great, I'll never use FP16 again.
Anyone baking?
>>101839929lol based. good celeb genning on sd15 is brutal. they can turn out so perfect. and then you deage them slowly, knowing that the real version has long started decaying. time is a bitch.>>101839995negligible.
baking rn
Why does Flux like to blur images if they have white backgrounds? Very weird quirk, seems like a training data problem, like maybe a disproportionate number of the TD images with white backgrounds were also low resolution
98% of instagram "photos" are more fake than failed sd images
fire up the collage maker, bakermaninclude my gen plz
>>101840033Default https://comfyanonymous.github.io/ComfyUI_examples/flux/ workflow on flux-dev at fp8_e4m3fn weight_dtype and fp16 clip is around 15 seconds.>>101840043>>101840072>>101840107Cheers.>>101840036On an anonymous image board? Get real.
Let me know when Flux can produce actual kino like this.Prompt was simply>Guts, berserk, closeup of bruised face, view from the side,(HunyuanDiT 1.0 base)
Interesting, it generated a line in the middle on this one.
>>101840208>view from the side>isn't actually a view from the side
>>101840212Did you prompt some physical media? It looks like a fold line on some poster.
>>101840223Here your Flux view from the side bro
Look what I found
I told it to write 4chan and Flux decided to turn racist wtf.
>>101840282>>101840282>>101840282
>>101840237I think it's implied in the prompt I'm using.>Drawn in pencil by Tsutomu Nihei.I've seen gens where the edges are distorted like you'd see on scans, but this is the first time I've seen the fold lines in a gen.Also, testing different resolutions again.
>>101840208Tried your prompt on flux....