Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101990155>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>GPU performancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>Maintain thread qualityhttps://rentry.org/debo>Related boards>>>/g/sdg>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
Blessed thread of frenship
...guys?
Always making the collage trend continues for me. Fuck I'm awesome.
>>101992825Post results
>>101992825What, it's only going to take an hour. Go make some tea.
With flux I don't even want to generate coomslop (and instead generate just slop)
>>101992829Very based fren
>>101992797thank you baker
>>101992847I'll coom to this in your place.
Most current LoRAs destroy Flux's gens and I'm bothered that nobody seems to notice of care.
>>101992847>and instead generate just slopsame
>>101992858Thanks fren>>101992873Extremely based
>>101992847It's not like you can generate better coomslop than you can with XL atm so
>>101992895Yeah. But in SDXL I don't think the military/sci-fi aesthetic can quite be nailed down like I can with flux so I am just generating what SDXL can't really do at the momentT-this is my last one I promise
>>101992895Depends if I like tasteful nudes or not.
>>101992917I'm checking Fluxtastic on civitai right now and I'm noticing pussies are deformed, and the model still ignores lots of lewd stuff.I enjoy generating late 2000s,/early 2010s amateur looking lewd selfies with XL much more right now.
if only you knew how much json chat is winning
>>101992974I don't gen porn at all. AI porn isn't a turn on for me.
Ok I lied but this goes hard
>>101992829How does it feel, anon?
>>101993004It's porn you can control. humans like control. And we will get even more control as time goes on, with better finetunes and full animations of movies
>>101993138Hey! I know her!
I’m seeing posts of people running a 3090 on a Thermaltake 700 PSU, but I have that one and it only has 8 and 6+2 PCIe power connectors. How are they supposed to run the 3090 if it requires 3 rails and 12 pins?
>>101993144I hate control. That's why I set my seed to randomize after every gen.
>>101993138Billy Idol?
>>101993119It's just nice to be noticed even if my GPU did all the work.
>>1019930042010 faproulette and “jailbait” style self shots are my thing, but with the advent of camwhoring that has been ruined too. That’s what I emulate with AI. I’m into nostalgia fapping.
The speed with which we're getting a whole ecosystem of flux accoutrements spooled up is really encouraging
>>101993223You are doing the work that pays the electricity bills, and you come up with the prompts, so it’s a partnership.
>>101993161who is she
>>101993256I don't think she would make a very good barista
again with the charcoal hands
>>101993191Well of course that's the fun part, but you still ask for things you would like to see, just the unexpected makes it extra fun.
>>101993243I just don't have that. I still have sex dreams about my wife.
>>101993266She's Morena Baccarin, most well known for the TV shows such as Firefly and V.
>>101993298ty friend
It can do nude just be creative with it
>>101993289>posted on 4chan
>>101993320my beautiful nubile daughters
>>101993243how do you even emulate something so specific? train it on the photos from that era?
>>101993323Been here for 15 years. We're here for life, Anon. Some of us get lucky and find someone just as autistic as ourselves.It helps that I'm not fat.
>>101993365You're on a porn website first and foremost.
Glass didnt distored background fucking usless ai
>>101993377I'm on a blue board. I have nothing against coomers, I'm just not one of them.
>>101993386His button is in the wrong place on his jacket
>>101993398or you're just a virtue signaling faggot"I don't like porn like everyone else on 4chan"
>use gguf model in comfy and the override/force to device script>put transformer to gpu 1>put text encoders to gpu 2works fine.but when I put a lora loader after the transformer ("unet") loader, it only uses the lora for one single image.what causes that?
>>101993417I like tasteful nudes, I already said. Calm down.
Changing lora parameters and reloading all the models is harsh, oof
>>101993436Nah, you wanted attention with your virtue signaling. "I don't like porn and I have sex with my wife". I'm sure she would love to see you browse 4chan.
>>101993429>put text encoders to gpu 2Always a catch, isn't there
>>101993317np babe
>>101992797Is flux 1 good at generating lolis?
>>101993346I use RawCharm - Amateur Photography and it does it well enough for me.
>>101993453Sounds like you're mad and alone. Sorry Anon. Enjoy your porn.
>>101993417or you are feeding a troll. >>101993429I am just getting back into using Comfy and apparently the detailer in the impact nodes won't do batching anymore. The same enshitification could be your issue too.
>>101993474Or maybe I'm married too and I don't bring it up unsolicited. And no, I don't browse 4chan in front of my wife either.
>>101993489>here's my kink>oh ok here's my kinkyou: REEEEEEEEEEEEEEEEEEEEEEE
>>101993346NTA but yes that's the planhttps://civitai.com/models/664143/sears-lingerie-ad-flux?modelVersionId=743250
>>101993455what do you mean?
People argue about the dumbest shit on this general it's unreal.
>>101993533You clearly haven't visited /sdg/ yet
>>101993533And what are you going to do about it? come at me bro!
>>101993543I was there from the start, I am glad the split happened.
>>101993533cool wizards
>>101993567sdg has been bleeding into ldg, sdg is getting slower and slower
>>101993574Thanks>>101993579I don't really mind, the only annoying thing is it sets off the schizos every time an avatarfag appears, and then they have a melty too, and the whole thing becomes retarded.
>>101993482That's weird too. Those nodes have always worked good for me.
>>101993612I mind because they're clearly high schoolers so their conversations are retarded by default and they don't go on Discord because they're already banned
>>101993633Why do they even put it in the OP then?
>>101993623it is by design apparentlyhttps://github.com/ltdrdata/ComfyUI-extension-tutorials/blob/Main/ComfyUI-Impact-Pack/tutorial/batching-detailer.mdCould be my poor memory or it could be king shit pooped out trash code again. I am past the point of expecting things to just work. >>101993579it was a raid last night. I wouldn't mind so much if their work was better.
>>101993657why do you think they use 4chan if they're actually using Discord?
It's cute how the stupid phrasing in this and the old OP "Use a VAE if your images look washed out" is still there, as if using a VAE is somehow optional.
>>101993663>I wouldn't mind so much if their work was better.I just don't really get it. They have literal infinity at their fingertips and create the same thing every day for 2 years over and over.
This is my wallpaper
>>101993691i never use vae on my pics
>>101993668some people aren't tribalistic
OK, so my PSU actually has two pcie power connectors (6 and 6+2), but the 3090 I got has 2x 8 pin slots. I can’t connect this any way right?
>>101993691Canny remember the last time anon asked "why do my images look washed out" kek
>>101993695this .. I mean let em do it, but its boring to look at
>>101993710Now it's my wallpaper
>>101993711Yes, you do. That's required to get an ordinary image.
>>101993657mental disease, literaly. schizo broke down because city posts here now. It's retarded, he seems to have started some sort of parasocial relationship with a bunch of anons that just ignore him.tldr: ignore the schizo, and any posts talking about the schizo, as that is likely a newfag or the schizo himself, let the other thread be forgotten.
>>101993736That's why I'm here and not there. That and the whole social club clique aspect of it.
>>101993750i ain't never use no stupid vae
>>101993721no, but you just need to buy an adapter>>101993765well I started off there so I sometime look over there but as the other anon said, the schizo vs avatarfag battle is just offputting, you post something there and either side suddenly includes you in their strange battle .. its so tiresome
>>101993724yeah, that too. it's complaints you'd see for SD1.5 models
>>101993795I actually ordered a new modular PSU. I just wondered if I could do without it. But I don’t think my current shitty 700W one will be stable anyway
>>101993774enjoys 128 res images huh
>>101993816700W is abit low if its intel+3090, on amd+3090 it might be okay
>>101993668https://www.youtube.com/watch?v=dS9PwndF2iI There was a reason this exploded on tiktok. The idea that people will do anything to be noticed is just getting stranger as the reason for the attention doesn't seem to matter. >>101993695>create the same thing every dayI am 1 girl genner so I can't really throw that stone, but I will anyways. I assume it is some mind spectrum thing. You can't do good work if you are trying to bend a single character or concept over a long period of time. The work keeps getting worse. Even if they took the original concept and worked on backgrounds it would more tolerable. >>101993736the issue was they would spam the fuck out of the thread.
>>101993822nice SD3 gen
>>101993795I don't have that much energy for that much hate. It's funny how emotional they get on both sides over nothing though. >>101993816>2024>700W PSUYou live on the edge.
>>101992847Based SFW sloppers.
>>101993843it's only mental illness because normal people at some point want to progress, okay so you want to gen 1girls, that's nice, but at some point you either start doing some variations, start training, etc, or you take it to the next level and make a visual novel or something. it's really not much different from someone who draws one specific thing every day for years
>>101993897It's kind of lol lmao, because you begin to realize guys actually find pretty much any non-fat woman pretty, like literally dudes just want one thing, and it's disgusting (gf on a diet).
>>101993870I use two 165W GPUs with a 650W Gold PSU. Seems stable enough.
>>101993854thanks
Can a friendly anon please point me in the direction of a good repo or tool for creating Lora’s these days? I was using kohya before but I imagine there is a better option now.
>>101993843>I am 1 girl genner so I can't really throw that stoneBut do you gen 1 girl with the same face and outfit over and over again every day? There's definitely a difference. >he says, while scrolling his miku folder
>>101993897related, the absolute lulz of guys getting older and getting into milfs and I'm like, yeah, the plan was you'd get married, and she'd turn into a milf right about now.
one thing I noticed about flux when genning hires close up portraits of women is it tends to make bugs bunny teeth, pic related
>>101993941I would be scared to do anything with less than 1000W. Genning is thirsty work.
>>101993930>>101993949Huh?Did you mean to reply to someone else?
>>101993970it running 100% stable with a 4090, Ryzen 5800X and 4 SSDs on 800W .. just be sure its a high quality PSU
>>101993956Gotta try and make it gen a unique looking woman. When it's just woman or something it tends to look samey, same chin, same mouth etc
>>101993970>scaredShould be okay if attended, if it locks up, turn it off and you know you have a problem (you don't know if it's power).The problems with ps was that some of them were actually basically loooow quality, despite having a fancy sounding brand. At least, if you get a cheaply made component, it should have a dumb name like Yuilitoi [new Amazon company].
>>101993970>I would be scared to do anything with less than 1000Wit's like you have no understanding of how electronics work
>>101994012>>101994004I always subscribe to the philosophy to go overkill on the PSU always. It will only ever use what it uses so there's no harm in a 1200W PSU
>>101994004Which cards?btw, any word on optimization of Flux for AMD? I'm trying to figure out why it's so much slower, and if that can be changed any. For one thing, I don't even know what the nature is of safetensors, how are the files handled, etc etc. idk, it just seems like those who understand the technical aspects are the least articulate.
>>101993970no need for that I thinkthat anon has an expected efficiency of 590W. the GPUs would demand a bit more than half that if they were both used at max load.a single 3090 would work too
>>101994010that is probably true, some of the gens where I had several decrepitations in it was fine, need to keep that in mind from now on
>>101994045(for example, AMD cpu SUUUCK at the CAS cpu operation, if you can reduce CAS, AMD speeds up, similar must be true for AMD gpu)
>>101993927>>101993946people are infinite and it is great. Real vs animated clothing is a fucking trip. Getting animated loras into cosplay has been very fun. You learn to hate things besides hands. Fuck capes. I keep impainting them into perfectly ironed abomination of nature. >>101994010the 90s and female super chins are a fucking real and weird thing>>101994026I feel that people not understanding if their area will face a spike or an undervoltage is often overlooked. I have burnt through 3 UPS in the last year. I am at the point that I am just buying beefy PSUs and hoping for the best. >>101993946we all have that one folder.
Hi I'm using comfy UI but I'd like to improve my generations. I feel they are kind of blurry and want to sharpen up the lines and give it more realistic look.are we sharing workflows here?Also I've been using comfyUI over Voldy's lately is there a better choice or is it just personal?
>>101993739
>>101993764I am not the schizo, I have been lurking in these threads since their inception but mostly in gallery mode. Who the fuck is the schizo? The guy who made a bunch of demon boys? Are they conspiring on Der Cord?
>>101994065atbox?
>>101994094>are we sharing workflows here?no make your own, it's intuitive af
>>101993769nice, got anymore?
>>101994045>Which cards?PCI-E cards? nothing but the 4090>>101994037ya it doesn't harm.. for a new build id go higher to, but its simple to look up the max wattage draws of your components and if you stay under 80% of your PSU it will run in optimal range which is true for my system thanks for the silly low TDP of the AMD 5xxx cpus>>101994113you dont want it, thats a schizo SD1.5 prompt I recycled to see what happens, its a wonder it turned out good
>>101994094download chaiNNer and power through a bunch of upscaling models. I find the 1.5x upscale, 1.25 downscale will fix a bunch of issues like this. After you get a combo you like convert it into a workflow. personal choice based on what kind of fuckery you want. Video card power plays into it too. Some Comfy users are running 4090s and will just upscale/inpaint/detail everything because they have the juice and it doesn't cost them a load of time. You can't one-click outside of Comfy, but user like myself don't think that the quest for one-click is a good one.
>>101994045AMD is slower than NVIDIA for anything related to SD/LLMs/AI.safetensors is just a file format (and a lib) that stores the model in a state dict {"some.model.section": some_tensors}, wrapped by a class.
>>101993463You must bew here
>>101994125Well im using this one with LORAs
>>101994094It really just depends how autistic you are. You can get a lot more control with comfy but it's really not "intuitive af" as >>101994125 suggests.If you understand working with nodes already then yeah it's simple but there's a shit ton of nodes and no documentation.It's mostly just trial and error.
>>101994181>you dont want itI need it
>>101994129nice
>>101994094>are we sharing workflows here?quite often. just look for catbox links in these threads or check civitai
>>101994251heck I don't care, but dont complain>masterpiece, best quality, (realistic:1.2), bust shot, 1girl in a black nazi uniform with tattoos and a plain ruined city in the back, short blonde hair, undercut hairstyle, blue eyes, iron collar, (cyberpunk:0.8), spread legs, (meta baroness:1.2), art by Carne Griffiths, blockades, tanksyou had to prompt like that on SD15>https://litter.catbox.moe/fydhxq.png>>101994280link above is one, prompt is baaad tho
>>101994181>if you stay under 80% of your PSU it will run in optimal rangeIsn't this just for Bronze gpus or in general?
>>101994251to make me, make me feel heated
>>101994280got it, I'm routing my LORAS this way
>>101994310a well thats abit difficult to explain, cause first it depends on where you are, or more precise do you use 110V or 220V inputs, (the second gives you better power efficiency) but by a rule of thumb a PSU gives you the most watts for your input at a range between 30% and 80% usage, beyond that ranges your its just bad (even if its titanium .. or 80% or what norm, that just means it wont fail you then) .. optimal is always like 50% usage so overpowered PSUs are actually cheaper on the longterm (I just had a 800W lying around and hadn't AI in thought when I build this so was thinking about ~400W usage as middle in thought), read some better tech articles about this if you wanna know morealso poor USAfags with your 110v .. its just bad
>>101994440install the comfyui custom node manager. then you can install whatever custom node you like. it's pretty much a "mod manager" but for comfy.use some lora loader stack to load multiple ones.clip skip setting usually doesn't change much.skip the sdxl refiner completely, instead just do more steps or pass it through I2I.
Does anyone know a method to merge two sets of .txt files? Say I have the files [txt1(a), txt2(a), txt3(a)...] and [txt1(b), txt2(b), txt3(b)], how can I easily batch concatenate the .txt files -- merging only txt1(a)+txt1(b), txt2(a)+txt2(b), and so on. Don't really want to do it one by one.
>>101993377>You're on a porn website first and foremost.Porn is not allowed on this board.
>>101993795>no, but you just need to buy an adapterWent to the store and got one. Currently posting from my new 1girl summoning setup.It's an AMD CPU, so 700 should be enough. I'll probably keep the new 850 PSU I ordered as a spare or in case this one is too weak.
>>101994590Unironically ask a LLM, they're kinda good at this.
>>101994628just because your zoomer brain can only get off by people double fisting horses doesn't mean this site isn't pornographic
>>101994635just tried that and it gave me a python script that worked. lol, ty
>>101994590One way I think is pretty smart is first get a list of the files into a text file. In your notes, you take note of what ls gets them in the correct order.Then you have another command to concatenate them as you wish. It would be a bash loop, read a filename from your textfile, then you cat it to something like output.txt
>>101994632lets goo! have fun with your 3090!
>>101994709The thing is so heavy it looks like it's going to bend in the PCI slot.
>>101994717didn't it come with a spacer? most 3090/4090 do
>>101994717get an anti sag bracket for it right meow
>>101994564cute
>>101994749>pointy chinflux loves pointy chins.
>>101994741esp since they are cheap. Mine was $5. It has a magnet in the bottom, that is weak, so I can't move my case far with it in it.And I just have a 6950, which is not as big, but just looking at it, it's like, uhhh... that's dumb.
>>101994759it sadly is on so many glam shots, so I guess also in their data .. what tag could one use to counteract it tho?
>>101994759thats how people look like anon, people have chins, double chins are not normal
>>101994883
>>101994953skill issue
>>101994953>>101994883other chins. some are pointy, yes.
>>101992871noooo dont play with the new toy right after it came outthis is udumbass
There was some update to comfy that made it run worse for some people with modern torch or similar. Was it fixed?
>>101994976pointy is default.
>>101995021I would hug you and hang out with you if I could anon. You are never really alone. And you have a community here even if we are all faceless and nameless. You don't need to lash out like this.
>>101995031if you make your girls grin like crazy sure
>>101993840i think people just forgot that eren's always been a massive lamer because it had been so longmakes perfect sense if you marathon the manga and keep in mind the characters' personalities
>>101995050thats nice of you but you ignored the point of my post which is that your confusion comes from a failure of empathy
>>101995050kek
>>101993138why are there human heads on the shelf
>>101995176he's d*bonuff said
>>101995142that's some fucked up depth of field there
>>101995176they asked too many questions
>101995187>nigbo
World of Warcraft ingame screenshot, UI and HUD visible, the main character is Miku Hatsune, she is wearing a wizard robe and is holding a magic staff. The background is Icecrown Citadel from World of Warcraft. The action bar at the bottom has elemental spell icons. The minimap at the top right is a map of Icecrown.
>>101995242
>>101995228ok this one is kinda cursed
>>101995198I wish it would just default to everything in focus.
>>101993769how did you make the hands look so nice
why does ModelSamplingFlux need the image size? It makes it so the model reloads every time i change sizes.
https://www.youtube.com/watch?v=ymNFyxvIdaMThis is how I feel lately with Flux etc
>>101995394
>>101995412lol nice
Any prompt wizard care to lead me in the right direction?I don't know what I am doing nor how to improve.I use CPU only so bruteforcing my way into proficiency is too time consuming.
>>101995426buy a 3090
>Support loading long clipl model with the CLIP loader node.could this be used on flux?
>>101995455At most I could buy a 3060
>>101995463Good enough, full finetuning and lora training with 8GB soon
>>101995477*6GB
>>101995394sick music videotimestop fetish?
>>101995484>>101995477Let's round that to 4 Gb.
>>101995426>I use CPUspotted your problem
Can flux do prompt weighting yet? If not then why? I though you just scaled the text embedding vectors then normalized them or some shit. Seems like that would be no different with flux.
>>1019954633060 is great budget option
>>101995440fixed the flag for you
>>101995498this one needed fixing to
>>101995560baste six toed man
>>101995523The math is just not there yet.I think weighing is working on the clip embedding since it does change the image but not by much.
>>101995484>>101995503Not sure about that desu but 8GB is confirmed.
What is a good SDXL inpainting model?>>101995412Classic.
>>101995603*2GB
I have the q8 gguf working. I am only finding a pretty slight improvement in speed, now that I am not hitting low vram like with flux dev.I wonder why it's only a small improvement.
Imagine training on just 2 GB. We're going to be so back.
>>101995610That's a bit pushing it but you never know
>>101995242It straight up genned a vagina as the building entrance in the background
>>101995650just saw the trailer for BAR seems cool
>>101995619>I wonder why it's only a small improvement.a small improvement in speed compared to fp8 you mean?
https://civitai.com/models/667442/sxz-rutkowski-flux-loraRutkowskibros we're SO back
now this is warcraft
>>101995720in 20 years you'll be able to plug in an LLM to generate realistic ERP circa 2004 for you
>>101995760In 20 years you may be able to generate a realistic real time MMORPG experience with simulated population.
>>101995760I just wanna raid and get gear, man
>>101995775yeah maybe if you have an enterprise level set of gpus
>>101995711>https://civitai.com/models/667442/sxz-rutkowski-flux-lorakek, fucking Greg Rutkowski>>101995661smol version army
>>101995698well, compared to dev. My speed is basically:fast to slowq8 gguffp16 devfp8 (using dev, or the checkpoint give the same speed)[I haven't tried the n* one]My guess is my ram is pretty fast. 64gb, 3600.I mean, I'll stick with q8, it's faster for some reason.6950xt.
>>101995760in 20 years I'll be dead
>>101995883not with that attitude!
How much more demanding is flux over SDXL? Can I run it with a 12gb 3060?
downloading flux1-dev-Q5_K_S.gguf, gonna see what it does in terms of speed.
>>101995897gens are slow on it if you do large resolutions but not unbearably slow. you can train too which again is slow.
>>101995880>I mean, I'll stick with q8, it's faster for some reason.and it gives better quality output than fp8 so yeah, you are right to stick to q8
>>101995897much. yes, run fp8 or shnell.
I wish I had an invincible army of cats
>>101995903If you have the original .safetensors you can quantize it yourself
>>101995760>20 yearsYou mean 5 years
cursed
Which one should I choose? I guess the loras we currently have on flux weren't finetuning the clip aswell right?
>>101995941yeah maybe
>>101995923>invincible army of catsi can beat them
wow paladin miku
When I try to load Alex Jones + Pepe lora ComfyUi crashes but when I go for Alex Jones + MJ lora it doesn't, hmm...
>>101996157It would be too powerful, anon.
>>101996169fair enough kek
>>101996153
>can gen this thing in 30 seconds>used to take 15 minutes with my RX 5700hahaha holy shit I'm literally shaking
>>101996157How's your ram look while it's loading those? You do monitor resources if you get crashes, right?
>woman holding spearwhy does it always gen it the wrong way? its always tip down like an anime pose
>>101996157Isn't Pepe lora ridiculously oversized?
>>101995987would
>>101996191did you try prompting where the point is pointing
>>101996185>Q8_0 1.4it/s>Q6_K 1.7it/sIs this because of the mixed precisions? Why is it slower if there's spare VRAM in both cases?
>>101996153Lewd.
>>101996190>How's your ram look while it's loading those? You do monitor resources if you get crashes, right?I have 56gb of ram, and no it's not overflowing it, it never goes over 30gb, but everytime it crashes I have some peak in the gpu
>>101996194>Isn't Pepe lora ridiculously oversized?it is, fucking 600 mb, I guess that comfyUi doesn't like "big" loras or something?https://civitai.com/models/644109/pepe-flux
>>101996237CumUI is deprecated tho
>>101996202yes that doesnt work, i think i am figuring it out though. what matters is the position of the hand, the thing being held is an afterthought
>>101996266can Forge pin T5 to the CPU yet?
>>101996266only on CumUI I can put the text encoder in my 2nd gpu so I'm stuck in there ;'(
>>101996278hmm interesting
>>101996268>Miley Cyrus>>101996287>Avril Lavignepic related>a girl that looks like a mix between Avril Lavigne and Miley Cyrus
>>101996266fax
>>101996220that's the processor, not the ram
>>101996266Once Forge implements Tonemap + DualGPU + Guidance Negative then maybe I'll consider leaving ComfyUi
>>101996315like I said, it never reach the limit of my 56gb ram setup, It never goes over 30
>>101996266Gottem.
>>101996315two 4090s? your dick must be huge
>>101996336You DO NOT have 56 GB of VRAMYou can't even take the correct screen shot
>>101996358sorry I meant 58GB
>>101996338I can't wait to buy two 5090sOr really cheap A6000swhichever comes first
does the p40 do better at bf16 than fp16? bf16 is a truncated fp32, right?
Keep getting this error when I try to generate with flux. What am I doing wrong?>UpdateI have updated both comfyui and python, the issue remains.
>>101996371These models do not use your CPU RAM, they use your GPU RAM. It needs to fit on your GRAPHICS CARD. I bet you can't even post your GPU model.
>>101996358>>101996371oh you mean VRAM, I have no issue with that either, I have 24gb and it's only using max 16gb during the two loras inference
>>101996153GOLDSHIRE
Open that oven door and get your very own fresh loaf of...>>101996391>>101996391>>101996391
>>101996393You already reserved 15 GB doing something. You need 24 GB free for a default Flux workflow. Even 12 GB workflows will run out if you're reserving that much. The other reason is you're attempting to generate a massive image.
>>101996387wrong loader, you need "Load Diffusion Model" node, not Load Chekpoint
>>101996413>You already reserved 15 GB doing something. You need 24 GB free for a default Flux workflow.no, look I have a second gpu, the first one is loading the flux model (15gb) and the second one is loading the text encoder (9gb)
>>101996420but he is loading a checkpoint that has everything included, not just the diffusion model
>>101996430Well it turns out computers are machines, they don't work on magic. That means if you're running out of VRAM you're doing something stupid.
>>101996420How does this node work? A few parts of this confuse me a little.>unet_name undefinedCan't change this>weight_dtypeUnsure what this is>Doesn't give a CLIP or VAEDo I need to download these? They aren't included with the model?
>>101996194but it's so worth it
>>101996461>That means if you're running out of VRAM you're doing something stupid.but I'm not doing anything stupid, the first gpu has 9gb to spare and the 2nd gpu has 3gb to spare, how about considering that maybe the Pepe lora is fucked somehow or ComfyUi has a bug? never went through your mind or something?
>>101996474you don't want that nodethey are included in your checkpoint
>>101996491it's a you problem, that means YOU did something
>>101996506go fix your shit Cumfy
>>101996474default = fp16 .. the model is like 21 GB in VRAMfp8 (both modes) halves that and makes it makes it way smaller not exactly half, since there is overhead, but like 14gb which makes it useable on cards that dont have 24GB VRAM like 3090/4090
>>101996493Going back to loading the checkpoint with the checkpoint node brings me back to the issue in >>101996387
>>1019963152x4090 gang represent
>>101996474unet_name, you have it in the wrong folderfollow the official comfy guide: https://comfyanonymous.github.io/ComfyUI_examples/flux/
>>101996194overbaked too, seeing how every pepe made with it has artifacts as its natural texture
>>101996535Thanks. I'll see if I can stop pestering people by following this guide. Damn I wish I saw this before, kek
>>101996185>>can gen this thing in 30 secondsWhich card, let me try your catbox.
>>101996220sorry, system memory, are you running low?
>>101996672what is system memory? I only know ram and vram >>101996393
>>101996672>>101996693if you mean the SDD memory then I have 60 gb of memory to spare
>>1019966553090https://files.catbox.moe/pwmulg.pngAlso, this thing stays sub 70ºC at all times. It's incredible (previous AMD card hovered 95).
>>101996730>Also, this thing stays sub 70ºC at all times. It's incredible (previous AMD card hovered 95).glad you took the Nvdia pill anon
>>101996730GOD I want to trade my 3070 ti up to a 3090 so fucking bad. Having 8gb of vram on a 30 series card is just humiliating.Maybe I'll start shopping around my local trades
>>101996730>ViT-L-14-BEST-smooth-GmP-ft.safetensorswhat's this?
>>101996824a better clip_l modelhttps://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/blob/main/ViT-L-14-BEST-smooth-GmP-ft.safetensors
>>101996824>>101996730(obvs already dling it)
>>101996797It's worth it. I felt it was expensive before buying it, but I don't even care right now. I can use it for VR, since my GPU was the bottleneck until now, for textgen, image gen and gaming in general. It's also pretty future-proof unless things change a whole lot>>101996824I haven't actually tested if it's better. I just downloaded it because anons told me to.
>>101996875How would I know if it's better?
>>101996884It's supposed to follow instructions better or something.Maybe do some tests and post results.>ipddm/sgm_uniform is slightly faster than deis/sgm_uniform and it looks virtually identic
>>101996901I'd even say it looks better.
>>101996875The issue is, everyone selling a GPU is doing it because they upgraded. So the idea of proposing a trade for my 3070 ti really looks like it's not going to happen.
>>101996884>How would I know if it's better?I added "blur" on the negative prompt and VIT L did a better job than clip_l, dunno if that's a definitive proof it's better but at least it's somethinghttps://imgsli.com/Mjg3OTU5
>>101996933It's obviously not the same, but I'm keeping my old RX 5700 XT. You never know. Over the years you can end up building a secondary PC just with spare parts from upgrades.
>>101996957I did this over 20 years, and when I finally had to change the motherboard I had a complete computer I sold to a friend really cheap.
>>101996730>>101996655>30 secondsVery nice, my 6950xt took 319 seconds.If I change a few things it will improve a bit.
>>101996901>>ipddm/sgm_uniform is slightly faster than deis/sgm_uniform and it looks virtually identichow many steps do you use anon?
>>101997033My condolences. Until yesterday, mine took 15 minutes, so at least yours is usable if you iterate using the preview image.>>10199707020, but I haven't tested it. I think I read somewhere that with ipddm you could go to 15 even.
>>101997033Also, where did it save it? I just saved the image from the workflow.
>>101997089ComfyUI/output? Or it should've asked you for a location./ldg/ don't die now that I got a new card
>>101997248New bread
there's a new thread already, make the move people>>101996391>>101996391>>101996391