Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101896239>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>GPU performancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>Maintain thread qualityhttps://rentry.org/debo>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>mfw
>>101900441Are we to expect another round of mass CP posting tomorrow evening?
SOTA GGUF quants in your neighborhoodhttps://github.com/city96/ComfyUI-GGUFhttps://huggingface.co/city96/FLUX.1-dev-ggufQ8_0 VS FP16 VS FP8: https://imgsli.com/Mjg3Nzkx/0/1Q4_0 VS FP16: https://imgsli.com/Mjg3Nzg3
the absolute state of you lot. it's a fucking shame.
>>101900433dropped this
>>101900553>1girl, standing
>>101900553/ldg/ is far better than /sdg/ dare I say
>>101900441You have no communityYou have a dead thread full of schizo and low skill promotersYour thread moves at a glaciers paceYou spend your free time spamming ai cp in this thread to copeYou larp as a coder and rage when anon post proof of you saying that you don't know how to code a year agoYou are aloneYou are a loserYou are the thread schizoYou have been kicked out of every ai space as well
>>101900565this, so much this!
>meltdown ensuesWe might get our dose of CP tonight!
>>101900483>cp>featuring 0 childrenhow
>>101900584Were you here earlier in the day?
why can't ai niggas just get along?
>>101900483Ah, so you want to bring schizo shit into this thread too? That's cool, I'm game. I'll schizo it up with you if you want.
>>101900517City, will you marry me?
>>101900433collage???
>>101900595They actually get along really well in places that don't have censorship.
>1girl, standing>>101900595hello fellow human! i, too, wonder why humans can't just get along. as i recall, we human beings have made a number of musical songs that ponder the same issue. and yet, huma--i mean we, remain divided. perplexing.
>>101900517Here's my comparison Q8_0 vs fp16 vs fp8https://imgsli.com/Mjg3ODEz/0/0
>>101900610>>101900556
>>101900605Not schizo actually
why does every flux 1girl look like a granny?
>>101900620How about Q4 comparison?
>>101900608No I'm plapping command-r-plus
>>101900620A Q4_0 vs NF4 comparison would be nice.
>>101900632most of the realistic flux gens look toasted to me, similar to high guidance on sdxl, exaggerated features and contrast/saturation around edges etc.
>>101900639>>101900644yeah I should do that, adding Q4_0, Q5_0 and nf4-v2, brb
>>101900441Good job squatting for an entire two years, too bad you're a laughing stock now.
>>101900656Awesome, thanks
>>101900517It kinda sucks from what I gathered last thread the VRAM requirements for running Q8_0 is just over the 16 GB limit. Q6_K could be very compelling to overcome that hurdle but I do understand it will take additional code so understandably, it will take some time if you decide to do it. I would also wonder if it would be possible to apply .imatrix and other tricks you can do with GGUF quantization to help the transformer and if EXL2 would work and speed things up from GGUF which again would need time to implement. A lot of possibilities opened up so thanks as always for thinking outside the box, City.
>>101900441why is he farting pink?
>>101900654this can be fixed by stuff like the realism lora and CFG, but people don't often do it
>>101900640Based.CR++ soon.
>>101900441catbox?
>>101900588Oh, no. Ignore me then, I thought we were talking about the guy posting toddler gens.
does flux just have all violence removed from it?I attempted to gen some guys impaled with spears and it's just refusing
>>101900696use grok, sigma elon made a based and redpilled imagen you just gotta pay up. povertyfags btfo
>>101900691I am so fucking glad we have an actual proper competitor to SD. Maybe it will make them pull their heads out of their asses and not produce dog shit for once, who knows.
>>101900691I’m talking about the guy posting the toddlers earlier today who got banned after every post. He was debo, aka horn boy.
>>101900714Oh, okay then once again I have to ask.>cp>featuring 0 childrenhow?
>>101900710can grok gen this?
mfw being the only retard posting gens in a imgen thread. turns out it DOES feel good to be gangster.
>>1019007190 children because the janny deleted them
>>101900721i hope not
>>101900719I bet you enjoyed those gens a lot, huh pedo?
>>101900710>>101900721Grok literally uses flux
>227 views Let's see if it can beat the original 4.6k
>>1019007340 children before they were deleted, too.>>101900738No I'm the guy who asked him to stop posting tods.
>>101900713Don't forget that Flux does still have restrictions like NSFW and other things, and in the long run, I do think stuff like the Open Model Initiative is more important to tackle those smaller minor issues. But yeah, thank goodness FLux turned out to be the actual SD3 model people in the community were looking for. I don't even know how Stability will come back from that, they promised to revise their release in the future but at this point, they should just scrap SD3 and move to do an SD3XL or something.
>>101900721what model are you using to gen these?
>>101900739lalalalala i can't hear you. elon is a GOD and we are his soldiers.
>>101900754finally a good gen
>>101900517Is it just me or is q8 too slow on 3090? Here is comparison between the two, top is q8_0 and bottom is fp8, I am using this workflow https://litter.catbox.moe/l80e3j.png
>LIterally every other day a new format of flux gets released to appease vramlets>LoRAs and LoRA training gets fucked each timemake up your fucking minds.
>>101900769come up with a dataset then
>FiNaLlY a GoOd GeNsays the nogen.you are a SLAVE, and you don't even know it
>>101900779What does that have to do with what I said?
>>101900620Neat, thanks. Q8_0 seems definitely closer to F16 which reflects what I was seeing, I guess there's more difference the more complex the prompts are.>tfw can't put lossless in the readme now>>101900665Yeah, this is day1 so there's definitely a lot to be figured out, figured better to release it than sit on it for 2 weeks.Q8_0 technically has a few extra tensors that could be quantized but I was seeing some quality degradation with those.K quants are C++ territory so will definitely take some work.I'm not sure how feasible imatrix is for an image model (wouldn't we need to get the actual unet running under llama.cpp to record those values?) but it'd be cool if we managed to bootleg it in lol.I haven't looked at EXL2 much because I went for the P40 maxxing meme and lost my FP16 privileges.>>101900760That does seem slow, especially considering this comparison: >>101900620Not sure what would cause it, only time I saw those speeds was when it was falling back to numpy for dequantization.
>>101900769Punctuated equilibrium. There will be an explosion of advancements when we get a new model with promise, but within a month or two people will have settled into their choices. In a few days, someone will make a flux anime finetune, then a NSFW one, and we'll have restarted the old cycle of NAI and F22/Pornmerge.
>>101900739can flux gen this?
>>101900800The WEG crossover nobody wanted or needed
>>101900605I'l still l be here retard schizo, try not to overdose on tranny meds before then.
>>101900787it sounds like you're complaining about not having a GPU and wanting other people to do the work for you
>>101900800you are a menace to society. but in a (((good))) way.
>>101900749sd1.5
>>101900791Tested again. The first gen is okay (speed alternates but still not as fast as 25s), but then it slows down on second gen to that speed
Schnell GGUF please...
>>101900844rookie numbers
Can't believe Elon is Fluxpilled like usEveryone got tired of OpenAI's bullshit related to their Dall-e 3 API and open source is the future
>>101900721needs blue hair, but that marge is exuding sara jay energy so I kinda don't entirely hate her lmao
>>101900791>K quants are C++ territory so will definitely take some work.What is stopping you from importing llama-cpp-python to do that instead of translating the code from C++?You are right imatrix needs some sort of hackery to work to get llama.cpp to record those values so yeah, pie in the sky at the moment.P40 maxxing is rough but yeah, there is potential in someone making EXL2 work for this too. Should get roughly the same quality output results with equivalent bpw quantizations but it should be much faster.
Weird, Q5 was giving me OOM on my 3060, then suddenly started working. Normally I can even run fp16 without OOM, it just goes into lowvram mode and slows down.
The mood at SAI must be dire.
the future's so bright i gotta wear shades
>>101900911Looks plastic
>>101900924oh sorry. here, have some more grease.
>>101900935better
>>101900911he looks like he's wearing his skin
>>101900895i have no idea what any of this means, but my most fervent hope is that you smarty pants niggas make the fancy new model work for vramlet plebs like me. you have your orders, now make it so.>>101900935based
>>101900902Here is fp16 of same gen, so quantization isn't to blame for messing up the lengthy text.
>>101900844>>101900854>>101900902Yeah I think the backend is either not moving the model correctly off the device, not clearing vram properly, or miscalculating the actual weights sizes (though it should just be using torch.numel like normal). Will try to figure out why it does that tomorrow. Cancelling the gen and resuming will fix it for some reason.>>101900895I think llama-cpp-python is just a wrapper, it can't actually quantize weights from torch tensors. gguf-py from the llama.cpp repo recently got a PR that allows it to do that in numpy + has the dequant logic I used as a base for my pytorch ones: https://github.com/ggerganov/llama.cpp/pull/8939
>>101900791>>101900844So I figured out the issue, it happens when I randomize the seed. So top is a random seed, while bottom is a fixed seed. Not sure if it's a bug in the code?
>>101900903how can I send my gens to them? I would like them to see the fruit of their labors.
>>101900784I really like this one
THREAD THEME: make a cat
>>101900993me too>>101901019instructions unclear. catgirl as usual.
>>101901019
>>101900857Elon is hosting a completely closed model though, but if Flux devs get too woke he will use Flux.dev as opposed to Pro.
>>101901054yeah I think of it as a last option if you don't have a GPU
>>101900517All right gens, here's all the quants compared with each other: https://imgsli.com/Mjg3ODI0
>>101901067One day I'll stop buying HDDs and car parts, and get a 4090.
>>101901078Interesting, the Q quants and nf4 each keep different details. Hard to pick a winner, but q4 might have a slight edge personally. Seems like Q6 might be the happy medium of quality and size.
>>101900975I fixes itself if you interrupt the first run and start again, think that's what's happening in your image as well.>>101901078>>101901123There is also Q4_1 and Q5_1 which aren't too difficult to implement, could do the later as a stopgap.
>>101901123>Hard to pick a winnerfp16cmon, everything else is a compromise
>no matter what I try with kohya I get > raise NotImplementedError(NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.reeeeeeeassuming its because I have 32gb ram and it gives up, no idea otherwise since the other anon got it to fit on his 12gb 3060 rip, back to sdxl loras until even further optimizations for training appear
>>101901142I mean for the lower quants obviously
>>101901147you pick the largest that will fit in your vramthere's no decision to be made, you take what you can get
>>101901142Q8_0 is really close though, much closer to fp16, that's my go to model right now, I don't need fp8 anymore
>>101901155You are wrong, there isn't a clear winner between 4 to 5 bits.
>>101901155>you pick the largest that will fit in your vram>there's no decision to be made, you take what you can getnot that simple, nf4-v2 has almost the same size as Q4_0 and Q4_0 is much better, same thing for fp8 vs Q8_0
>>101900696Something I noticed about all latest models. No stabs, no middle fingers. They follow similar censorships. Of course MJ and Dalle can still do them no problem, goes to show that even at the top of the line due to censorship the quality of the data used for open source models is still bad. No stabs means you can't gen anything like pic rel which does suck.
>>101901078interesting that they all got the text wrong, or was there an accidental typo in the prompt
>>101901183no, it's just that flux cannot do this specific text somehow, can't work at 100% I guess
>>101901078>nf4-v2, Q4_0 and Q5_0 messed up the skin colorI wonder at what point it'll get it right, maybe at Q5_K_M, we'll see
>>101901144oh, just looked on Kohya's issues repo and it turns out his script only supports the FP16 weights. well, fuck, does that mean loras are only going to work with fp16?
>>101901210In the LLM world it's usually said that Q6_K "feels" the same as Q8/FP16, so that might be the one.
flux can make the absolute weirdest underwear
>>101901078It's impressive how much closer Q8_0 is to the real deal, fp8 feels like a shit quant in comparaison, and that's the one we were playing with for this whole time
>>101901193I wonder, would the text work if you placed it earlier in the prompt?
>>101901271fishing tackle lingerie, nothing weird about that
>>101901271well you got baited didnt you
>>101901271i own that and wear that. subscribe to my OF to see it all, tee hee. (ooc: jannies, this is satire)>>101901279ooh, me likey. catbox? it needs more blatantly obvious occultism, but it's got potential. i mean, you really gotta beat these dumbass niggas heads in with the symbolism. like, bash it into their pathetic brains.
>>101901329yeah nice but let's see more of her, like that, obviously
So who's who
>>101901336i am me
hello, focus your attention on me. i like it.
I too am that guy.
we all KNEW illuminati was white as white can be, but some of you pretended he was brown because you are always coping
>>101901344don't jerk me around
>>101901371HYPOTHESIS: YOU ARE NOT THE REAL QUAKKACONCLUSION: PUBLIC SHAME
>>101901359if this isn't a genre, it needs to be
>>101901359did she pee the jar full?
>>101901385Yum>>101901427It's honey.
>>101901437>It's honey.gross
>>101901447she pees honey
Richfags fear the CPUmaxxer.
>>101901476Use schnell, you only need 4 steps
>>101901493he's a man he doesn't need schnell
cuz i'm a nightmare dressed like a daydream.
THREAD THEME: 1girl farting
>>101901534you stupid brown bastard. the real thread theme is being WHITE
>>101901527nnightmare milkers squish squish HONK HONK
>>101901563your corpse will be white once i strangle you to death you stupid fucker, do the needful and start farting
>>101901579neat
>>101901579this is cool
>>101901520i still think the elf in a jar was art
>>101901595moar
>>101901595that's my jar
>>101901595farts into the jar and seals it shut
>>101901595None of this is art. Pick up a pencil
>>101901606the owls are not what they seem>>101901625this is cool too
>>101901595>$5 per lad
So I come from the old oobabooga days where there were a billion little syntax tricks in the promptwhat happened to all that shit now that everyone's using comfyui?
>>101901657oobabooga is text gen, do you mean automatic1111?
>>101901664yeah that lol it's been a while
>>101901595>pay up>dump a load>immediately hand over 5 more dollars>immediately bust again>immediately hand over 5 more dollars, much to everyone's dismay>she looks up at you in shock and disgust>still going an hour later
>>101901670Are you looking for /lmg/?
>>101901595
>>101901628>Pick up a pencilthe thing you use in school? what for?
need a good pinup for inspiration
>>101901697To learn to draw you talentless faggot.AI is a fad that's gonna go the way of NFTs and Crypto. Artists will win.
>>101901703
>>101901628ok
>>101901704then why are you crying about it here, baitanon? just watch it all crash down and make some art of it to put on Tumblr
>>101901493I have no need for such a crutch.
>>101901704>Artists will win.you will be winners among each other only, like wine tasters
>>101901714more
>>101901704>Artists will wincope
>>101901726how many cores do you have?
>>101901704Too obvious.>>101900517When can we stop fucking around with quants and model formats and just pick on and get LoRAs working on it? It's really frustrating seeing a week of work flushed down the drain because someone found a slightly better format for the base model.
shut the fuck up about the biggest cope of all: loras
>>101901742The onus is on you to explain why LoRAs are cope.What's cope about a lightweight and easily trained model for a very specific task?
>>101901734I dunno if it's really flushed down the drain, maybe the existing loras can simply be converted/quantized in the same manner to produce new ones that work with other model types.
>>101901704
>>101901754i simply do not require them
>>101901758can you make it rain matrix code?I want to learn how
>>101901756I hope so. It's getting silly having gone through like 4 variations of the same model in different formats in a week.
>>101901754nta but they're made entirely redundant if it's already in the model
>>101901761You will eventually. There will come a time where you need a specific thing that does not exist in the model.>>101901767This dumb, you cant sit on your ass and wait for a full fine tune with your hyper specific concept added.
>>101901779ive been doing this since you HAD to use the command line, before the outputs looked like anythingyou do not need loras
>>101901763
>>101901787retard
>>101901787Yeah. You do. Stop being obstinate. I'm not going spend hours and hours fine tuning a model when I can just whip out a LoRA in an hour or two. It's evidenced by the glut of endless niche concept LoRAs on civitai.
>>101901787Please create Taylor Swift with a hairy bush pissing onto an icecream cone without using any LoRAs. Post catbox. You have 10 minutes to complete this task, or you are a faggot.
>loras finally work on nf4>it immediately becomes obsolete because new quants
>>101901819kekuse a porn merge
>>101901823I still think the new quants look really really promising so it's worth pursuing. Just really frustrating having to see it go back to square one again.
>>101901823Shit is moving too fast, so I haven't created any LoRAs yet. This is what I tell myself to feel secure about my lack of ability.Seriously though if someone has a Kohya script (for gui) I can build off please post it.
>>101901378it was meant to be a cat...
>>101901829Tick tock
>>101901823I hope loras will work on gguf though
>>101901823>loras finally work on nf4Really?
>>101901855no
bad news fellas
>>101901823Yeah, nf4 fucking sucks compared to Q4_0 and they have the same size, that's funny when you know that the nf4 guy claimed that it's better than fucking fp8 top kek
>>101901855Yeah, in forgehttps://github.com/lllyasviel/stable-diffusion-webui-forge/commit/cb889470ba33722a89c3f625f972a795504abdc6What it does is it makes a new model with a patched-in lora(s), and it does it every single time you generate an image. Even if you re-gen with all the same loras, it will still make a new model so it will absolutely rape your ssd.those nudity loras are trash btw (at least for now) https://files.catbox.moe/biyel1.png
>>101901885Why does it look like 1.5 booba
>>101901885>Yeah, in forgeI sleep.
>>101901867What's the bad news?
>>101900517Will you also make it lora compatible?
>>101901904I forgot I was trying to give him a head the size of a walnut and it kept putting walnuts on his head
pretty sure I had a Trapper Keeper like this in school
We're never going to get lora training support using an optimized model, are we? :(
>>101901938what are they looking at?
>>101901941It was all looking good until the sub 24gb's started screeching
>>101901949each other (they are in love)
>>101901961I just want to make loras without spending $1k on a used GPU, man...
>>101901941the image gen space is 99% jeets and jeets cannot afford a gpu with over 10gb vram, so someone will definitely figure it out do not worry
>>101901961the lora training on sub 24GB cards was looking good until the sub 24GB card havers started screeching?are you retarded?
>>101901787>boomer is clueless about tech
wish i could use grok without a stupid x account
>>101901941An Anon trained a Flux LoRA on his 12gb 3060 earlier. Only took 25 hours.
>>101902000Just make a throwaway, but do you have to pay for Grok?
>>101902002That's exactly the problem
>>101902023don't worry, 20 years from now a $100 phone will have many times the GPU power of a 3060
>>101902028great for the people still alive by then
>>101902028Alright, alright, I'll get the rope ready
>>101902030should have eaten your veggies idiot
>>101902012yea x premium
>>101902023I'm gonna try train one on my work laptop that has a 4060 because I am the "AI guy"I couldn't convince them to pay for the 4080 version :(
>>101902042this image gave me some kind of phobia
>>101902041It's worthless then. Any musketeers paying for grok? Can it into porn?
>>101902056worms crawling up your backside?I think it's worse for them than you
>>101902047start taking 3x longer to do anything they ask whatsoever, it is only fair
>>101902002how did he do it?
>>101902047Godspeed anon.. I'm guessing I'll end up figuring out how to make the --save --resume commands work and sucking up the 40 hour 12gb training, but I hate myself for it already
>>101902067In all fairness I have access to essentially infinite A100s on AWS for work purposes so that's not a good excuse. But I'm not racking up company AWS costs to train my Belle Delphine LoRA.
>>101902091Pure CLI. It was either last thread or the one before, you can find it. He gave a bunch of info.
>>101902096pussy
>>101902109May me $140k a year and I will.
>>101902096Tell them I hacked you, then you share it when it's donewin win
>>101902107ah, here it is>>101898699>>101898925
>>101902124>the notorious hacker known as 4chan
>>101902023No, people like you are the problem. You demand everyone conform to your pathetically small vram and throw a wrench in progress by demanding the model format change each day to contort to your increasingly unreasonable demands to fit SOTA models on your old busted shit GPU.
>>101902096man I know the feel, I could probably finetune the fuck out of flux for porn if I could get away with using university resources for it but that would be a very bad idea
>>101902148spends 1k to gen shitty 1girls
>>101900517Q8_0 is impressive, it gives me good images even at 20 stepshttps://imgsli.com/Mjg3ODY1
>>101902165
>>1019021651k is not enough for a real card
>>101902148>i got mine fuck youYou know some of these models took 40 GB+ to inference some years ago? Everyone benefits from optimizations. But since you are such a purist, go ahead and turn off xformers or split attention. Go for fp32 while at it.
>>101902190>i got mine fuck youcorrect. Stop fucking up progress because you can't be enthused enough to buy and enthusiast card.
>>101902148The fuck is your problem, if you don't want to use optmized quant models, you do you, let vramlets have their fun
>>101902202>missed the pointthat is why you just ignore retards
>>101902202>Stop fucking up progressi guess optimization isn't progress then, just gotta keep on bumping up that parameter count. got a problem? just throw more compute at it! accelerate! accelerate! accelerate!
>>101902148You act like the optimized models are gonna break into your hard drive and steal the full model from you
>>10190217330 steps feels like the sweet spot
>>101902212NTA but Anon, if I ignored every retarded thing that was said on this board I might as well not even come here.
Pedoposter i got ban becouse of you
FLUXDEV NF4 JUST FLEW OVER MY HOUSE AND DROPPED CHEMICALS ON ME
>>101902148I don't give a shit about your faggotry I just want to make loras within a reasonable time frame, plain n simple
>>101902224true true, you can see it there https://reddit.com/r/StableDiffusion/comments/1er3wt7/if_you_want_a_good_compromise_between_quality_and/
>>101902067Yeah he should walk like in slow motion too. Sorry boss gpu too slow
>>101902222OK but you see how flux is being optimised to fuck already though right? It's already happening, I'm not really sure what you're complaining about.
>>101902258>Flux is being optimized and this hurts me! this is a troll, right?
holy crap the tech is moving fastlast time I used SD1.4 a year ago and everything nude was pretty shit, now I just installed it again with cyberrealistic 1.5 and the first results are way better than anything I've generated beforenow I know why imgfap has so many AI stuff in the galleries now, some of them I most likely didn't even recognize as suchit really is over for amateur models making only pictures
>>101902281you should see the current SDXL finetunes
>>101902279I think you've misinterpreted my post.
Is it possible to use any version of flux on an RX 5700 XT (8 GB) and 64 GB+ RAM?Also, how do you quickly toggle upscaling on and off, and move images from preview to inpaint and back quickly? I'm used to sdui and I find bypassing and changing connections on comfy a bit cumbersome for my workflow
>>101902289Bro dont do this to yourself
>10190214880% sure this is the schizo switching tactics because jannys banned enough of his pedo posting proxies desu
>>101902289I recommend you sell it and buy an NVidia GPU, even if it has the same VRAM. You will get significantly better performance.Sucks to shill but it's simply true. AMD fucked around with the AI performance of their cards and now they are finding out.
>>101902258???
people told me I should get the 3060ti because it's better for the same pricethey ain't laughin now
>>101902310What do you mean ???. Flux already runs on 8GB cards after only 2 weeks, and further improvements are coming.
>>101902222>>101902324I'm confused how you think my response was complaining about flux when it was pointing out how angry the message from anon sounded.
>>101902321better than what?
>>101902288>>101902324I think you're getting confused replies because of the post you quoted
>>101901732In most of those cases, nobody was replaced. They just upgraded their tools and optimized the way they worked.All this pitching cattle against each other is so tiring.
Look, I don't care about your quants that let you run this shit on 500mb of vram or whatever. Stop breaking LoRAs
>>101902287So for that I need to download the main SD XL model and then another checkpoint and use that as refiner? Or would that be the Lora stuff?
>>101902297>>101902306I know. This card only works with --force-fp32, that's why I asked. There's a lot I need to spend money on before I can get a new GPU, so I guess I'll just wait.
>>101902374You just download the finetuned checkpoint, no one uses the refiner, it was dropped from consideration right away.
>>101902380And rocm5.2 exclusively by spoofing the card's vendor string. It's pure pain.
Hello frens, here is a comparison of the quants. Unfortunately I can't test Q8 without getting OOM, even though I can run fp16. Might be fixed soon when the github gets updated.In any case, Q5 clearly adheres pretty closely to fp8 here, much better than the lower quants. It's also the largest quant that fits in my 3060. So it looks like an ideal quant for us 3060 vramlets.
>>101902383Ah thanks. Do you have any checkpoints you recommend? It's hard to find a decent one for coom shit when all the NSFW images aren't shown.
>>101902393speed?
>>101902393>Unfortunately I can't test Q8 without getting OOM, even though I can run fp16.that's weird, Q8 "only" asks for 16gb wheras fp16 asks for 24
>>101902402For porn mostly used RealVisXL with the NSFW POV lora and then used PonyXL (the autismmix version, more stable style quality)Haven't touched XL in a while tho, I hear NatVis looks very good
>>101902393Q6 probably fits into 12gb, i am sure there will be more quants, this is just the start. it's also interesting how the difference is much more subtle at cfg 1 when compared to >>101901078
>>101902403All about the same until you get down to something small enough to fit all in VRAM (Q5). Then it's about 4 s/it for me, rather than the 6 I usually get.>>101902410I'm guessing something about this relatively experimental implementation of gguf quants isn't properly offloaded to RAM like the regular model files are
have LLM style quants been possible all this time?were they never needed until now
>>101902440>I'm guessing something about this relatively experimental implementation of gguf quants isn't properly offloaded to RAM like the regular model files areyou should use the --highvram flag and force the model loader to only go to your gpu with thishttps://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/even if you don't have multiple gpus, that can be used to force the loader to go on your only gpu (cuda:0)
>>101902436>Q6 probably fits into 12gb, i am sure there will be more quantsthere will, the _K quants like Q5_K_M, those are the really optimized quants used on LLMs (no one use the _0 anymore)
>>101902432So you used a model combined with a lora and another model? My only experience is working with SD 1.4 and then upscaling it with Gigapixel.
>>101902432sorry should've clarified I'm not looking for HC stuff
>>101902445ye, anything "AI" is pytorch, except google shit but who cares that's never open source. Cept Gemma which is fine
>>101902445The comfy gguf guy's github says they were shit for previous models but work well for transformer/DiT based models like flux
>>101902306AMD knows exactly what they're doing. The ceos are first cousins and AMD exists so NVIDIA can point and say "look, see, we aren't a monopoly"
>>101902491For RealVisXL yes since it can do nudes but doesn't know the raunchier stuff.>>101902500then grabbing any popular XL finetune should be enough, you can go on civitai.com, filter by SDXL 1.0 and sort by most downloads in the last month.
>>101902445it only works for Flux because that one is a transformers model (like the LLMs)
>>101902511and intel? did they get the same memo
>>101902393When text is right at such a low quant (which is impressive) there isn't such thing as a baseline. Q4 is best gen.
>>101902516No, they came in after the fact and are falling flat on their face.
>>101900517Do you have a specific reason on why you decided to use those clip models instead of the "regular ones" for flux? https://github.com/city96/ComfyUI-GGUF?tab=readme-ov-file
>>101902547those are the "regular" ones but with the original names from OpenAI and Google respectively
>>101902520hands are fucked on q4, it will probably be the same with things like guns.
>>101902511The modern world is a big fat lie all of it innit
>>101902331>>101902328Oh right, I was the one who was confused.
>>101902547see https://github.com/black-forest-labs/flux/blob/c23ae247225daba30fbd56058d247cc1b1fc20a3/src/flux/util.py#L129
>>101902563i was gonna but i got hit with a five minute timeout so i'm going to bed instead
>>101902557I just have to say you're confusing a lot of people this this github image, I'd suggest you to photoshop that part and put the "regular" names kekhttps://www.reddit.com/r/StableDiffusion/comments/1eslcg0/comment/li7kp0x/?utm_source=share&utm_medium=web2x&context=3https://reddit.com/r/StableDiffusion/comments/1eso216/comment/li7n6re/?context=3
>>101902562Hands are kinda fucked in all of them desu. Q4 and nf4 both mess up the bottom text a bit though, the rest don't.
>>101902590OK I will bake
I went ahead and tried anyway. I downloaded nf4 and the loader module, but when I try to queue anything it says>module 'torch' has no attribute 'float8_e4m3fn'Like I said, this card (RX 5700 XT) only runs on rocm 5.2 afaik, so I'm stuck on pytorch 1.31.1. That's probably why I'm getting this error, isn't it?
Well ok, here it is...>>101902610>>101902610>>101902610
>>101902595Nta but if a bunch of Reddit retards can't figure out that the two models are the same and can't do the bare basic task of seeing if their current t5 and clip models work, why should I care?
>>101902625kek, fair enough
>>101902514>grabbing any popular XL finetune should be enoughcool, thank you anon
>>101902597Nah, read the prompt closely. Q4 and nf4 at least tried.
>>101902393damn, nf4 really sucks, can't believe forge dev claimed that it's "better than fp8", what a fucking moronhttps://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981
>>101902685I look like that
>>101902795cutting out gigabytes of weight and still thinking it can perform as well, really stupid
Anybody experience lots of blur in Flux dev? Most images come out like this.
>>101903148you often have blur if you use cfg > 1 without dynamicThresholding
>>101902878isn't this the entire point of the quants? Q8 cuts a lot and performs about the same as fp16. i think that is just a bad quant.
>>101903159Thank you so much, that may have been it! Messed with those settings, and although they still look iffy, the blur is gone
Guys, what if I told you I managed to remove the blur on photos with the ToneMap node?https://imgsli.com/Mjg3OTE5
>>101903291you're welcome anon, if you want to mess up with CFG > 1 without having this blur artifact, use DynamicThresholding https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/
>>101903345Cheers, anon!
>>101903330>>101903330Huh, interesting - this reduced high-cfg artifacts for me as well. >CFG: 6>Guidance Neg: 10>Adaptive Threshold: 0.992>bokeh, blur in negative Still need to tune in the values but the difference is immediately obvious, while still having the benefit of high guidance/cfg. https://imgsli.com/Mjg3OTI1
>Nothing to smile about
>>101903505yeah, feels like tonemap is a better version than dynamicthreshold for humans
If you want a better clip_l, someone improved that shit:https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/blob/main/ViT-L-14-BEST-smooth-GmP-ft.safetensors
>>101904568Wtf? It made my blur removal even better
>>101903330If anyone's interested, I made a long ass tutorial to remove the blur of the pictureshttps://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/https://imgsli.com/Mjg3OTU4
