Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101951994>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
fuck off debo
comic: pepe enjoys his day with miku gens
>>101954396
THREAD THEME: go jerk off before genning
comic. Pepe is dressed as a samurai and is holding a katana. he is standing in front of Mount Fuji in Japan.
>>101954445How dare you assume the female form is only for jerking off, get out of here you sexist pig.
>>101954451that blade does not fit in that sheath
>>101954461That sheaf is there to whip your ass when you say something annoying.
>>101954460i was walking about the pepe spam. hnnng.
>>101954451That's a scimitar. Is your Pepe a filthy turk roach?
>>101954468harder daddy
>>101954474oh...my bad.
Has anyone ever successfully recreated reality from a doodle?
so close...
>>101954590img2img it at lower denoise
Retrained my lora, actually worth publishing now
https://civitai.com/models/657191/softserve-anime-flux?modelVersionId=735293Impressive
>>101954633nah shit's still smudgy, train more
>>101954633you need more steps for the face
>>101954682That's the point anon...
>>101954607
skinny fit girl with big bob wearing a tight t-shirt?
>>101954737this is not the place for that>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>>101954737Yeah sure, give me 10 mins
one piece style lora but with miku hatsune as the prompt:
>>101954818woah mama catbox?
>>101954841yeah sure https://files.catbox.moe/y0j5rs.png
>>101954797I tried with a basic prompt of nami's clothes. neat
I get black images as outputs with flux gguf. What am I doing wrong?
>>101954696better?
>>101954922you need at least double the nodes
>>101954934workflow?
>>101954737>>101954748I swear, those boards are filled with 80 IQ subhumans begging for AI porn.
>>101954922Your positive and negative is the same.
>>101954922Change your cfg to 1 instead of 8.
>>101954922https://comfyanonymous.github.io/ComfyUI_examples/flux/
https://civitai.com/models/655408/flux-monochromemanga?modelVersionId=733261Nice
What's the current meta for outpainting?
>>101954782>>101954737This is for you Anon.>>101955016and for you
>>101955072you call that big?
>>101955087It's big for a skinny girl.Ok give me 10 more mins.
>>101955072thank you friend
>Load up Q8 Flux>Load a LoRA on top>Watch as the model loads, 13.5GB.>It moves the sampler>14, 15, 18, 23.5gb of vram being used>huh?What's causing this?
>>101955182T5?
Loras are so good now that you cripple yourself for not using them. Thinking of training a few myself.
>>101955192T5
>>101955209and you have just the one GPU? which you are pushing T5 into?
>>101955220Says right there, cuda0.
>>101955016/aco/ anons are sharing interesting flux gens
>>101955087How about now?>>101955112No problem
>>101955224that doesn't tell me how many GPUs you have but you do understand you're putting the diffusion model AND T5 on the GPU, of course VRAM goes upyou can run T5 very fast on the cpu and keep it in RAM at all times
>>101955230much better, almost didn't notice the fucked up leg, almost
>>101955230what did you use, Bombshell or Big Boobs FLUX?
>>101955249Fine. I'll just deal with the 20 second wait every time I need to alter my prompt. No biggie.
>>101955283it's time to upgrade CPU and RAM, anon, that 7700k has run its courseprompt processing takes a couple of seconds on a ryzen 7700
>>101955230better, but still BIGGER
>>101955293I'd rather just shove my old 2070 super in my mobo on riser cables than deal with buying a new mobo.
>>101955316to be clear if you're not OOMing you don't need to put T5 on the CPUbut you should have figured out what was taking the VRAM on your own just looking at your workflow, dummy
>>101955340Why can't we just quantize the t5? It should respond just as well as the model.
>>101955312>>101955310
>>101955259Same.>>101955272This onehttps://civitai.com/models/649980/cleavagehelper-flux>>101955310Ok give me 10 mins....
ff tactics lora, miku as a wizard, I guess the diff jobs are instance prompts.
Normally I suggest tagging images in LoRAs is often more optional than not... but I really should have tagged the nude images.
>>101955393>Normally I suggest tagging images in LoRAs is often more optional than notsounds retarded, anon
>>101955371can you give a huge titted demon a different expression like the kevin macalisters, as if she's internally saying "oh fuck my tits are impractically huge"
>>101955389
I still get a black image with flux gguf
>>101955436you need, like, triple the nodes
>>101955436anon what the fuck are you doing
>>101955402Okay. I will be a good boy about tagging in the future.It's just so annoying>>101955444>>101955436Loading the exact same clip model twice probably isn't helping either.
>>101955436clip_name 2 should be the clip l file
>>101955459loading the clip
Is the OOM bug when loading a lora model on ComfyUi fixed yet?
>>101955434and a knight.
can Forge put T5 on the CPU yet?
>>101955459>>101955463i only have the flux clip in my clip folder, what do i need to download
>>101955459>dual cliploader>loads clip twiceI dunno anon that makes sense to me
>>101955371>https://civitai.com/models/649980/cleavagehelper-flux>every caption during training had "a woman's cleavage" at the start.brainless lora makers
>>101954721touching the blade is taboo on many levels, not the least since your filthy greasy fat fingers will rust the steel
>>101955552>two left hands
>>101955559she's an android, so?
>>101955415Ok...give me 10 mins...
>>101955415NTA but I don't think the tits are big enough.
>>101955499https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensorsBut there is a finetuned version that is better albeit 4x the size so use that instead unless your hardware is unusually weak.https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/blob/main/ViT-L-14-BEST-smooth-GmP-ft.safetensors
>>101955310
>>101955589>But there is a finetuned version that is better albeit 4x the size so use that instead unless your hardware is unusually weak.it only loads the text encoder, it doesn't take more memory
>>101955567there, she got a tuneup
>She has her hands on her cheeks in an expression reminiscent of the cover of the home alone movies with her mouth open in surprise
>>101955590>tfw will never have giant booba minigts-gf ;__;
>>101955627I've never even bothered to use references in my prompts, I just assumed they would misinterpret
So, has anyone tested the new K quants?
>Single GPU 24gb fine tuning unironically coming very soon.Who excited?What fine tunes will you make?https://github.com/kohya-ss/sd-scripts/pull/1374
>>101955710https://github.com/kohya-ss/sd-scripts/pull/1374#issuecomment-2295166778>Additionally, I have introduced a technique to save memory in gradient checkpointing by offloading gradients to the CPU. The combination of these two approaches should enable training at a reasonable speed even with 24GB VRAM.Oh shit that's real
Some retard gave me 100 buzz on flux for posting an image. He wasted 10 cents because I posted a pic for free
>>101955733Time to change your name to queen bee with all that buzz you got there.
how much better is q6 vs q5 and q4?
>>1019557471bit better than 5, and 2bits better than 4
>>101955742Damn right. But why the fuck do people do this? Do they get off on it? Is it a sexual thing?
>>101955762you value money differently
>>101955762ultimately everything is
>>101955635I didn't know what minigts meant but my search ended up here...so thanks anyway. :ohttps://www.threads.net/@katherineleanor002/post/C8A-3UZg5vc
neat, it knows game shows.
>>101955747If you've ever worked with LLMs you'd find that Q4 is about as low as you want to go before a model starts taking a massive dive in quality. everything above Q4 right up until full precision is incremental. The difference between Q8 and Q4 is significant but not as big as the drop in quality between Q4 and Q2. You might not even notice it upgrading to Q5 from Q4
It's retarded how easy it was to train a Flux LoRA of my wifes face what the fuck. Way easier than it was for 1.5 or XL. Even when I got it going on those models it was still somewhat off. This one is straight up her.
>>101955762Conspiracy time: Civit probably randomly gives out buzz to incentivize people who usually do not have it to try and spend it and maybe purchase more in future to use more of the sites features.
>>101955710Me, also they said 12gb vram might be possible too with some train half and swap to other half technique.
>>101955795>It's retarded how easy it was to train a Flux LoRAYeah, turns out those 12 billion parameters are pretty useful for capturing likenesses and even very fine thing the human eye doesn't pick up on.
>>101955804someone here claimed to have produced something with 12gb, but he went silent after he got it working so it's probably junk
>>101955800I can see the guy who gave it to me. It's just some guy.
no loras for schnell, sad
Why is there a flux Q4_K_S but no Q4_K_M?
>>101955812Yeah no doubt. I used to post gens using the wife LoRA because it looked enough unlike her that we were comfortable sharing them here. These ones, I cannot, it would instantly doxx us.
>>101955844I don't think anyone really gives a fuck about schnell desu. I'd rather wait 40 seconds for a fantastic image than repeatedly hit roll on boring slop. The only people who really advocated for schnell were drama queens with a soapbox to stand on screeching about schnell being their model of choice because of the license. But everyone just kind of ignored them and went to dev anyway. They'll follow eventually.
>>101955865The only ones who advocated for schnell were vramlets who now have Q4 which is compatible with all dev LoRAs
>>101955880schnell uses the exact same amount of vram
waiting for a flux anime finetune before retraining my retro anime lora
>>101955865>everyone just kind of ignored them and went to dev anywayAnd I am glad for it.I was so worried the community would adopt schnell as the default due to the vram issues back at the flux release.
>>101955851You need to mix different types for half of the attention and feedforward tensor layers where your type, Q4_K_M, half of them is using Q6_K and half are using Q4_K and that support hasn't been added to the custom node. If you use all the same type, you end up with the S quants which is all the support that was added in.
>>101955885Yes but it gens 5x faster because it does it in 4 steps.
>Everyone celebrates Flux finally being on par and even beating saas models>Not even a week later ELON MUSK of all people released an AI that shits all over flux and is completely uncensored
>>101955898which doesn't change the amount of vram it requires>>101955880>The only ones who advocated for schnell were vramlets
>>101955800Also, I almost have enough buzz to train a flux LoRA on site, 113 buzz off.
>>101955911Elon uses Flux pro
did they finetune soul into flux yet?
>>101955911bait
>>101955913Yeah but it overflows to RAM. If it takes 30s per step to gen would you rather receive an image in 2 minutes or 20? This is the difference.
>>101955415Here you go, I couldn't get her to look like a demon much. >>101955582Noice.
>>101955892Is the problem that we have no perplexity equivalent making it so that we can't know which tensor have more influence on the result?
>>101955938if it fallbacks to RAM they won't be using it anyway
>>101955922Oh man, you're so close. Maybe if you engage more with the site by posting more images someone might give you more buzz! And hey, even if you don't, you can just buy more buzz then make your LoRA so people can then engage with it! Maybe you'll even get more Buzz for your LoRA.What a steal!
>>101955940did you copy my post verbatim as the prompt?
>>101955911
>>101955953The more you buy the more you save!
>>101955960Just partly:A demon succubus very big boobs with a huge cleavage and a suprised and shocked open mouth expression, very slim waist, wearing a tight party dress, a thought bubble that says ''oh fuck my tits are impractically huge'
>>101955940Gross. I hate women with small tits. >>101955953I know it's jewish trickery but if I post more seemingly Anti-Russian memes I might just get enough.
>>101955975buzz, kofi, clout, tesco clubcard points, it's all the same shit
how many layers of irony is this
>>101955975Sorry about the small tits, desu I would need to make a lora to get bigger than small tits, maybe train on tits big enough to destroy Godzilla.
>>101956004Let's count themHand drawn image created by AIDrawing of Emma WatsonSigned by Emma WatsonIt's a photograph of a drawing4.
>>101954922set cfg to 1, load clip_l.safetensors
>>101956016https://civitai.com/models/658359?modelVersionId=736641
Why does ComfyUI disable spellcheck in text boxes now by defaultI'm too ESL for this
>>101956017and every step is illegal because of the deepfake thing, so double it
>>101956037Your prompt is probably shit anyway, just ask an LLM to make them more verbose.
>>101956050your mom is shit
>>101955975>Hello sir good norming I am training Taylor Swift LoRA *Posts flux image of Taylor Swift* but I need buzz to do the needfuland then collect your buzz.Here I even made a PFP for you to really sell the persona.
is flux working well on a1111 or forge yet?
I guess caricatures need a degree of (dare I say) artistry to know which features to exaggerate
>>101956076indian luke smith
>>101956086Forge: Yes?A1111: Anyone still use this shit?
flux armpit hair?
>>101956140I use a lora for armpit hair (and pubic) but it's only for asian woman.
is it normal that everytime I change a lora I got an unload + reload of my model on ComfyUi?
>>101956140yeah
>>101956140https://civitai.com/models/649568/asian-female-pubic-hair-and-armpit-hair-for-flux1-dev
>>101956155I think it's a bug but yeah, it seems to be "normal"
>>101956155Yes. It loads it then releases it.
>>101956157sniiifffffffff
>>101956165>>101956166that's not convenient at all, I hope this is a bug and will be fixed by Comfy or something, it wasn't as much as a pain on A1111 for example
Which one of you made this?https://civitai.com/models/658897/fluxpup-thin?modelVersionId=737259
>>101956155it has to reapply the changes to the weights so yeah, it's more noticeable because Flux is much larger than we used before
>>101956175When you say A1111 did you mean SDXL etc?Because even in Forge it seems to load up the lora every time you change a lora, which is pretty annoying but I guess it's needed
>>101956183>0 downloads and reviewsIt was you, wasn't it?
do you accept good ol' sd 1.5 pics here?
>>101956200Absolutely not, go fuck yourself.
>>101956195No I swear! please no, I just refreshed the new loras and saw it, I've been refreshing the page all day to look at new loras, please believe me it wasn't me.
>>101956184So basically you better have your fun with your lora and find the good strength value or else you have to wait 2 mn for any new modifications? Now I have even more reasons to wait for a finetune that will add more concept into it >Because even in Forge it seems to load up the lora every time you change a lora, which is pretty annoying but I guess it's neededYeah, only the lora, on comfyUi it unload the whole model before reloading it again
>>101956200yeah, as long as they are good any model is fine
>>101956200yeah if it's good. Some of my fav gens are still from 1.5 day, it felt like it had a bit more soul.
>>101956204Alright, I'm off to kill myself
>>101956218no don't listen to that creep any model is allowed
>>101956218Wait at least post an image first so we can judge it.
>>101956207>or else you have to wait 2 mn for any new modificationsget more and/or faster RAM
>>101956184>it has to reapply the changes to the weights so yeahI think I see what you mean, basically it has to delete the "modified" model and summon the original flux again to apply new weights is that it?
>>101955209>>101955220>>101955249I have the same input, why should I put T5 to CPU? And how do you do that with dualcliploader?
>>101954365So is this shit NVIDIA only? I have an Intel Arc A750...
>>101956232too late, i can't stop the bleedinggoodbye
>>101956255>too late, i can't stop the bleeding>goodbyeBefore you die, let me try this as a prompt in flux, Stay with us my Pc is slow.
>Went from Flux is impossible to train to flux is actually incredibly easy to train and the results are fantastic within two weeks.
>>101955911Grok seems to use a mix between schnell and dev
What is the main difference when prompting in a combined clip_l and t5xxl CLIP node for flux?I've never split my prompts up before in this way so I don't know what sort of parts of the promp is best suited for each section, hapl?>>101956248Force/Set CLIP device node
I downloaded the risitas lora, but I can't add other characters with him, how do you solve that? like when you use a character lora you can't add anything more to the picture? https://civitai.com/models/657967/el-risitas-juan-joya-borja?modelVersionId=736191
>>101956271Seems to be dev but with more training, so slightly better, but the community will make dev way better eventually.
>>101956271I guess it could seem that way to absolute retards but it isn't.
>>101956283>Seems to be dev but with more training, so slightly betterLike a Dev but without the guidance distillation, AKA Flux Pro.
>>101956267Alright. I'm barely hanging in there, but I'll try.
>>101956276>What is the main difference when prompting in a combined clip_l and t5xxl CLIP node for flux?I tried a shit ton of combinaison "with, without, both" and I didn't find any significant difference so I don't bother with it at all
are there any "must have" flux finetunes yet? Most of the civitai examples don't look amazing
>>101956283>dev but with more trainingLike... a professionally trained version of dev. Maybe they call it dev pro?
>>101956076I did the needful. Turns out if you like pictures you get free buzz. Now I have enough to gen a LoRA on site.
>>101956276Where is that in the comfyui basic menu? Is there any way to search in that menu that opens with right click, I installed too many plugins
>>101956307Probably won't be for a few weeks. Single GPU training likely out in a day or so, but good fine tunes will require data be properly tagged and prepared to get good results.
>>101955842I literally posted about it last thread you faggot
>>101956155For me, it's even worse, have to unload every 2-3 gens when using Lora, for some reasons it keeps using more and more vram.
>>101956316
do the proper tools to make a good LoRA exist? I have two 24gb cards. Heard you might to add natural language captions to all your data?
>>101956296oh, the pain...
>>101956344>for some reasons it keeps using more and more vram.there's some memory leak in your side, fortunately it's not my case, I think you should make an issue about that
To train a lora in kohya, do I use the base model I know the checkpoint uses or the actual checkpoint I want to train the lora for?Like, if I use a pony variant do I use pony base or the actual variant model?
>>101956342I posted about it this threadwe're like brothers
>>101956366you're right, let's celebrate
>>101956365Usually training for the original, SDXL model is enough, if you train on a finetune the results of the LoRA might only be acceptable for that particular tune. The exception is pony, like you said. That being said, you can still get good results training on fine tunes, just expect the unexpected. Literally nobody knows what they are doing.
I think a higher CFG helps a lot when you want to modify a lora's character style or some shit>a man ElRisitas in the style of a 50's comic bookhttps://imgsli.com/Mjg4NzU4
>>101956348kohya ssjoy captioner run locally
>>101956348I tried florence2 last night and it's garbage>adding a pop of color to the sceneextremely repetitive and cursory
>>101956399this gives me such a weird uncanny valley feeling like he's mixed with a pug and those weird tall bugs from men in black, wearing the cartoon flesh of a humansend help
>>101956296>>101956255sorry for the delay, here's the resultYou still with me?Hello?.....
>>101956420kek
>>101956422thanks, you made my last moments funbye
>>101956422second result
>>101956414>>101956348https://desuarchive.org/g/thread/101948110/#q101950141
>>101956394I'll try with the finetuned variant first then, than you anon.
>>101956442llama3.1 is VLM? didn't know that
>>101956184>it has to reapply the changes to the weights so yeahWhat happens when 2 loars are loaded and they share the same weights? one will have the priority over the other? And that's why you can't spam multiple loras at the same time I guess?
>>101956434No problem, see you in another life, brother.
I don't know why it should reload the model when you change the lora's strength on ComfyUi, why not keeping the "modified" one because we're just changing the same modified weights at the end, that's weird
>>101956280Lower the weight or use a different LoRA or wording around it, it's all dependent on the LoRA.
prompt: hookereh?
>>101956345Weird, with it set on cpu it doubles the s/it time, on cuda:0 it's 10 seconds faster
>>101956296>Alright. I'm barely hanging in there, but I'll try.This as a prompt gives interesting results.I should try more random phrases.
>>101956280>Two people - A and B doing...>A is dressed in...>B is...Anon, have you still not learned anything?
>>101956618>weird, my GPU is faster than my cpu Uhhh, anon.. you feeling okay...?
>>101956658Wait I mean this as a prompt.>>101956255
>>101956668Without the force set clip node I got a similar time, this node basically increases gen time by using the gpu? Why was it not set like that by default?
>>1019566891) it depends on your specs2) comfy is a talentless hack
>>101956532Even better than that, why reloading the model at all? When you apply a lora, why doesn't it keep on memory the old weights that got modified and when you put them back
>>101956666>ElRisitas shaking hands with Donald TrumpI think I put the lora strength too high, everything becomes a Risitas
why do some of you gen intentionally ugly and gross subject matter? I don't even mean aesthetically unpleasant, I mean legitimately disgusting stuff like >>101956721is it just to see that the model is capable? why would you want to look at this
gottem
I don't understand the excitement over having Pepe loras in flux. SDXL has always made sufficient Pepes
>>101956666prompting doesn't save you from loras that like to make every subject the same, and when you lower the strength to avoid that it makes the likeness worse
>>101956760I'm gonna use this comment as a prompt, be back in 10 mins with the result.
>>101956775Everything is better in flux and if it isn't it needs more Loras. Sdxl is archaic slop to be discarded, the future is now
what's flux?
>>101956796>it gens kamala harris
>>101956823some crypto bullshit
How do I run joy caption locally? Do I just take the script from the hugging face and plug the appropriate models and file paths in the script?
>>101956823Flux inhibitor
>>101956414Use this.https://github.com/StartHua/Comfyui_CXH_joy_caption/tree/main
>>101956796genned gore for me
>>101956835:>>101956442
>>101956852I've got something else working, but thanks
>>101956852>Chineseimmediately closed the tab
https://reddit.com/r/StableDiffusion/comments/1ev68la/cartoon_character_comparison/Holy fuck I never expected Schnell to be so bad
I’m using an AMD RX 5700 XT (8 GB) barely struggling to use Flux. I’ve found a 50% discount on a 3060 with 12 GB.I assume this will be good to use a quantized version of dev. But what s/it numbers am I to expect?Also, will this be a decent upgrade in general for gaming and VR?
>Genning in comfy UI>23.6GB of VRAM>Genning sample images during LoRA training >19GB of VRAMHm?
>>101956889Lmao schnell has sovl
>>101956899I know money is an issue, but I don't know people would pussyfoot around with a u new GPU for this kind of stuff and not aim for 24GB.
>>10195689912GB is not enough for dev
>>101956760>>101956796The result.>>101956825oh you naughty boy!>>101956854Lovely.
>>101956899Everywhere I look the two cards seem to be equivalent, the 3060 being slightly better of course. I think it’s dumb to spend money to barely improve.
>>101956915say it with me now: comfy. is. a. talentless. hack.
>>101956928>>101956943this, you go all the way or you go home
>>101956899~4s/it
>>101956943yes it is>>101956947So basically Flux detected the negative energy from that guy's comment and created a image that represented that.
>>101956953she isn't
>>101956966xe/xir is
>>101956947it's kind of cute I'd call him snorgle and feed him thunder stones until he digivolved
>>101956957>>101956963How you get 4s/it with dev? How many steps?
>>101956979it's zoomed in, it's smaller than the head of a pin
>>101956966pardon me, I forgot he's actually quite talented at unbridled autistic melties and anal sex (receiving). carry on
>>101956955Fuck. That means upgrading my PSU as well. Thanks for the spoon feeding frens
>>101956979That's nice of you. :)Here's another result but I'll catbox it incase a mod doesn't like this.
What the proper way to make empty clip for negative prompt when using flux? I use like the example with another clip text encode that is empty but feel like it's processing it for no reason.
sdxl + flux's text encoder when??
>>101956993even better, he won't take up any pocket space
>>101957005forgot the catboxhttps://files.catbox.moe/t9xr78.png
>>101957023I'm scared to click, chat what is it
>>101957047>Hello? Domino's? Someone took a shit on my pizza and spelled the word "dead beef" with it.
>>101957062>worse, they cut it in quadrants
>>101957062Kek I was just thinking this exact thing when I saw the thumbnail
that one .gguf lobotomy of flux, even if i get that shit running on my gtx 1080, is it even compatible with loras? Iike if i did a character lora train right now on civitai could that shit work just fine?
>>101957069that's okay, I'm trying to cut down on how many slices I eat.
>>101957079>that one .gguf lobotomy of fluxThere are way more than one.
>>101956989Not doing anything special, running on forge with Q4_0 and T5 in fp16. Using nf4 it can even do ~3.5s/it.
>>101957069my sides
>>101957089>Using nf4This should not be a consideration. NF4 is garbage.
>>101957052Kentaro Miura's last work.
>>101957087>
>>101957108I wouldn't say it's trash, but there's a reason I'm using Q4_0 after all.
>>101957047if you imagine it to be chocolate brownie, then that's a very tasty pizza
>>101957069I'm laughing so hard because I read this as if you'd be complaining they cut the shit into quadrants to spell out the words.. ah, I should really sleep
What's going on here? I only got 10s once when changing device from cpu to cuda:0, then it only increases gen time
>>101957108>This should not be a consideration. NF4 is garbage.this, there's no reason to use nf4 when Q4_0 exists
>>101957079LoRAs work with the gguf quant models without anything extra. Just use the normal LoRA mode.For the specific one you're talking about I have no idea, I have the proper quant models.
Can you gen something that captures the feeling of the no (You)s blues?
>>101957131dominos did actually do a pizza base with chocolate fondant instead of passatathe first bite was okay, then I got nauseous very quickly
>>101957108>help me mommy the optimized model makes images that are different from the rest and im scared>different is bad right???NPC
>>101957141someone already did many threads ago:
OK civit let's see how you do.
>>101957141>Pepe lora: feels bad man
>>101957159Hey that was me, kek
So, I see you can run fluxdev on Intel a770 now. Is this new or am I behind the times?I seem to recall recently someone here saying it wasn't possible.https://old.reddit.com/r/comfyui/comments/1ev7ym8/howto_running_flux1_dev_on_a770_forge_comfyui/
>>101957173lmao that was an objectively good gen
>>101957187Flux advancements happen hourly, the future is beautiful
I think the gummy worms was too much
>>101957210>eat ze bugs
does chatgpt api have memory option without eu restriction shit?
>>101957205>Computer, draw Audrey Hepburn as a Kardashian, extra plastic sheen
>>101957205cute
>>101957192It was an objectively bad gen. It was full of monstrosities but his face was definitely a winner
god DAMN people are going HAM on the Flux loras. Most of them are shit!>all those pixel art related loras that look no better than what's out there for even SD1.5the worst part is seeing how much these loras lobotomize flux's anatomy understanding..>>101957257absolute failgem
>>101957158Is this weak bait? NF4 produces results that are wildly different from all other quants. That's a really bad sign.
>>101957210you guys are making it very difficult for me to eat breakfast today :(
>>101957262If it looks good who cares, end result is what counts
>>101957158No one got fooled by your lies lllyasviel
best gpt frontend?
>>101957322chatgpt.com
>>101957322maybe ask in /lmg/ instead
>>101957322>>101956888
>>101956899save money for 4060 ti 16gb or used 3090
>>101957270sorry, wholesome pizza henceforth
>>101957283If the end result is unpredictable and strays from what parent model produces it's a strong sign that something has gone very wrong and might be lacking in areas that aren't immediately obvious.
>>101957358>puffy pizza>cheesy date hook lineimpressive, very nice.
Fresh load at the ready...>>101957363>>101957363>>101957363
Finally I managed to add miku with him kek
>>101957358Cute
>>101957379Based.
>>101956442on this do I have to run the model locally via say ooba and joycaption can see it?If I run say mistral large that will use up all my vram over both GPUs, does joycaption need extra or is it sending requests to the open LLM
>Replicate API FLUX dev = 0.03 per image. >1 hr 4090 rental: ~0.40/hr>~13 replicate images = 4090 rent 1 hr>ASSUMPTION: 30 seconds gen time per Image>120 images/ 4090 hrI want to generate around 40K images using flux dev, and the easiest way would be using an api provider but doing the calculations, its a complete scam, am I missing something?
>>101955765SEGA's biggest mistake.
>>101956979>I'd call him snorgle and feed him thunder stones until he digivolvedI put this as a prompt.
>>101955842I made this lora with 3060 12GB, here's how the character is supposed to look like https://danbooru.donmai.us/posts/4258364
>>101958087you better not be trying to create YET ANOTHER synthetic dataset or else
So I should use WD14 to train a lora for a pony model right?Does it matter that much?
