Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>102003576>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
>touohoutard avatarfag attentiontranny still forcing his slop into OP collagesembarrassing, literally mentally ill
>>102006777>not a single of those animal crossing/y2k gens made the collagewhat the fuck are you doing? last thread had tons of good gens.
>>102006801>touohoutard avatarfagHe generally doesn't post any text along with his slop so many people don't notice it as much, but I know when an avatar has broken from their containment. I saw the quokka come in here once but he left when people weren't responding to him.
>>102006811I posted those late, I made the doom miku tho
>Hey guys I'm just gonna post a link to my patreon and some random pictures of my LoRA in your pull discussion on github.
give me some random flux lora ideas that i'll probably never train. i was thinking of a pikmin enemies one.
>>102006777checked, and thank you baker
>>102006846tuktuks from various countries
>>102006846Gather a bunch of world of warcraft classic screenshots and tag them appropriately.Flux already knows WoW. Help it understand the prompt better.
>>102006846a 1997 isuzu hombre
>>1020068468 bit nintendo games
>>102006846bbig boob lora
>>102006863>Gigantic windows on first floor facing public beachYeah nah, no way I could ever relax in that house knowing any number of crackheads could be watching me from the darkness of the beach outside.
>>102006846futa and sweaty butthole lora
now that I am bored with bored women in FLUX, thanks to the lora .. what to prompt?
>>102006889futa I already seen on civitai .. actually like 3 or something
>>102006777imagine licking ran's sweaty abs and then brushing her tails after a workout...
>>102006882just pile lots of crack and fentanyl in the woods and they'll swarm it like zombies and leave you alone
>>102006901a lora for each of the /sdg/ avatarfags then
>>102006901yet no sweaty butthole lora, what has our society come to?
what's the difference between CFG and guidance? I'm confused
>>102006918That is some 90's nostalgia shit.
>>102006918she's going to vore the moon and poop it out
>>102006929>That is some 90's nostalgia shit.that's the goal yeah, I fucking love the y2k erahttps://civitai.com/models/667307
Blessed thread of frenship
>>102006926they trained Flux on CFG 1 so they had to find something else to improve the prompt adherance, and that's how the distilled guidance was born, but fuck that I'm CFGmaxxing with Dynamic Threshold or Tonemap, CFG still rules to this day
Forge gives me 3.4s/it vs Comfy's 2.4s/ittried both FP16 (guessing Forge is casting to FP8 automatically) and the Q8 ggufIs Forge a joke?
>>102006846frutiger aero, y2k has one
What is the least I need to get a lora to work with dev? I have the nodes connected as usual, but am I missing something with the required Guidance?
Seaking of loras, I got a technical question. Is it pointless to use LoraLoader with CLIP strength? I see very little impact when varying that from 0-1, is LoraModelLoaderOnly enough?
>>102006954Flux Pro uses CFGFlux Dev was guidance distilled to not use CFG but instead have conditioning on the Guidance value so it learned what different CFG levels look likeFlux Schnell was guidance AND step distilled
>>102006975>Flux Dev was guidance distilled to not use CFG but instead have conditioning on the Guidance value so it learned what different CFG levels look likeso there's a way to bring the regular CFG back to flux dev with a finetune?
gguf flux isn't actually faster or is it? or am i doing something wrong?
>>102006975Was it ever concluded if it was possible to un distill flux dev? Not like, restore it to pro state, but uncrack it a bit.
>>102006983>so there's a way to bring the regular CFG back to flux dev with a finetune?no, till now DynamicTreshholding with CFG normalizing to 1 is the only way to get access to CFG on dev, a finetune wont change that.
>>102006986GGUF is just better for quantizing and (don't quote me on this) plays with system ram a little better?
>>102006990>no, till now DynamicTreshholding with CFG normalizing to 1 is the only way to get access to CFG on dev,tonemap also works but yeah what a bummer, I thought it was possible to "un-distill" a model with some more training
Anyone else notice teams that use the T5 encoder get really nervous and dismissive when you talk about finetuning it?I wonder why that is.
>>102006990>a finetune wont change that.If your finetune stops conditioning on guidance or you fix the value you are effectively training it without guidance distillation>>102006999it is
>>102007011>If your finetune stops conditioning on guidance or you fix the value you are effectively training it without guidance distillation>it isthen it is possible to bring the CFG back if we undistill the model during the finetuning?
>>102007019NTA, but I believe that is the idea.
>>102007009it's a waste of resources and would mean you can't have one T5 shared with all finetunes>>102007019you can't "undistill" back to what it was but you can make it better for CFGCFG is something that happens during image gen, not trainingduring training what you need is conditioning dropout and Flux Dev's distillation process might still have done it
>>102007068damn that looked good until i opened it...
>>102007072>you can't "undistill" back to what it was but you can make it better for CFG>during training what you need is conditioning dropout and Flux Dev's distillation process might still have done itat least it's something, if it can handle higher CFG better that's already a win, and it won't be so fried on high cfg + DynamicThresholding/Tonemap
>>102007068so they can make kale and spinach look like a burger, but they can't make mealworms and beetles look like a burger
>>102007087what happens when the tide comes in
U vill eat ze bug
>>102007092Can't explain that.
Miku Hatsune is in the game Animal Crossing. She is talking to Goku from the anime Dragonball Z. In the background is a comfy looking house with a sign that says "Kame House". The image is in the style of the game Animal Crossing.
>>102007092>let's keep looking honey
>>102007103so you managed to make it work without the animal crossing lora?
>>102007110it works better with the lora because it was trained with images like it, the base flux even knows what ingame WoW looks like.
512 tokens is a lot, they really haven't missed when they released Flux, the only mistakes they made can be easily fixed
>>102007104make me pineapple pizza plus spiders
>>102007136different tokenizer
>>102007141true, but I don't believe the difference is that big maybe 5-10% off the GPT4 tokenizer in terms of tokens total count
>>102006991i don't understand what is going on anymore. why is flux schnell smaller in size than flux dev? for example the q5 variants??
what happens when you try an XL lora in flux?
>>102007118even the most popular mmorpg ever? amazing
>>102007170you get an error, they are not compatible
>>102007182you're joking but it doesn't know the main character of GTA5, one of the most popular games ever
>>102007166>why is flux schnell smaller in size than flux dev?It shouldn't be.
does flux know old school runescape?
>>102007170this anon is kinda right >>102007189 .. it won't crash tho, comfy just looks at the lora and tries to applies it and sees its in the wrong format, you get alot of >lora key not loaded: ...in the console, and then its like no lora is loaded at all
I asked for an ar15 .. it gave me an akm .. well I guess its fine to
>>102006777lucky thread
Running into all sorts of OOM issues I wasn't before recently. Switched back to the fp8 that was running fine at the beginning but it's still happening. Gens fine at 1.09s/it for a few gens then shits itself and goes as slow as 1m/it. Anyone else experienced this?4080 btw.I would be interested in seeing some of your workflows if you're running the same card and not getting this.
>>102007227Schnell not making thick outlines :( . What proomt u used ?
>>102007283try the q8 model + gguf loader, seems to work better/less memory issues
>>102007319That's what I was using and where the issues began, and why I switched back to F8.
>>102007334this has been working pretty well for me, same cardbut f8 is fine too
>>102007345Hmm, you're using clip_l, I am using ViT-L, maybe it's that?
>>102007358that might be it, just use the ones in the comfy guidehttps://comfyanonymous.github.io/ComfyUI_examples/flux/
>>102007283need more info, and maybe screenshot of workflow. lora involved? you offloading clip to ram?
>>102007363I mostly am. A few tweaks for uptional upscaling and adding a LoRA and that's it.
>>102007390see >>102007391>you offloading clip to ram?I guess so. 32GB RAM is full up, that's the issue.
>>102007391looks good, I still need to learn stuff like adding a section for inpaint stuff if I want to, been using forge for quick edits to test so far.>>102007396nice, could be the cover of a ps2 game, lora or just prompts?
>>102007403>lora or just prompts?it's a lora yeahhttps://civitai.com/models/667307>Y2K style cover art with a 3d character of Hatsune Miku, with her iconic turquoise twin-tails flowing in the wind, performs an impressive kickflip on her skateboard while soaring over the majestic Grand Canyon. The sun is setting in the background, casting a golden glow over the vast, rugged landscape. Miku’s expression is one of pure exhilaration as she defies gravity, her outfit a blend of futuristic and streetwear style, complete with neon accents that make her stand out against the natural beauty of the canyon. Y2K style text at the bottom: "STYLISH"
>>102007391I've seen that workflow before. ehh try w/o lora. loras add nasty overhead to model - i got this node that can force clip into ram
>>102007425Yeah I did and it's the same issue, but I want the LoRA. And it was working fine with the LoRA until just a few days ago.I feel like in my tweaking I connected up something wrong somewhere and fucked the whole thing up by putting the wrong node into the wrong hole.
>>102007416it's a cool aesthetic, reminds me of playstation/dreamcast game boxes
>>102007458w..wrong hole?!try the clip offload thing https://gist.github.com/Sunderbrazeput in base custom nodes folder, force clip to cpu, NOT on cuda aka your gpu. worth a shot
https://www.youtube.com/watch?v=9NkkZJHova4
holy shit I actually got a name tag on this one.
>>102007506I tried forcing clip to cpu but I gave a 10400 and was slow as fuck when I change the prompt.But, I didn't have this script, so I will try it.
>>102007545
>>102007545Which lora is this?
>>102007312I used a Lora: >https://civitai.com/models/668807/simple-ink-illustrations?modelVersionId=748712
where's the finetunes mansly
>>102007559I know but I am OOMing with the dev-fp8 model and some loras and I got 24GB. can't update comfy because #2666 fucked various custom nodes I rely on so GGUF is a no-go for me>>102007526BRO lol
>>102007605two more weeks
>>102007589none, vanilla fluxWorld of Warcraft ingame screenshot, UI and HUD visible, the main character is Miku Hatsune, she is wearing a her default outfit and holding a large greatsword. The background is Hellfire Peninsula from World of Warcraft. The action bar at the bottom has sword icons. The minimap at the top right is a map of Outland.
>>102007628and there is a typo, but it still works!
THREAD THEME: create a gen of what you think the average /ldg/ anon looks like
>>102007570sex
>>102007657I feel nice today. You are all needy catboxers.
>>102007657
>>102006897Very Nice. Catbox please. I need a fix for fine arts style in Flux pronto.
>>102007687erotic male
https://www.youtube.com/watch?v=UYahPVL3FL0
>>102007705this is male
>>102007748>the nooseayy lmao
Flux base model is too huge and has garbage data...I don't want ugly old politicians nor WoW data in it.
>>102006846Asterix
>>102007769the point is the base model can do almost anything. just use loras for specifics it can't do, or certain styles.
How do I use the gguf t5 clip? Keeps throwing up an error
>>102007769>Flux base model is too hugethat's the thing, it's big so it can eat a lot of shit, everyone will be happy if everything is included
>>102007769t. vramlet
Is it possible to train flux loras using 1024x1024 pictures on a 16gb card, or do i have to downscale them to 512?
https://www.youtube.com/watch?v=ZhFAj8-Wfx8
cyberjews
>>102007867>muh mikuu
>Trained 7 LoRAs so far and they've all been of my personal porn collectionThe LoRAs... they're just so good, it's like I can just keep making more porn from the porn I already have.Like planting porn seeds on a porn farm to harvest more porn.
>>102007911make a Dafne Keen pony lora
You know what a good lora style would be? the one mimicing the FFX cutscenes:https://www.youtube.com/watch?v=QgW-UC9tcU4
>>102007545>>102007570nice
>>102007911>so as I pray, infinite porn works
>>102007657not physically but mentally
>>102007802Update comfy, they pushed an update yesterday
>>102007911>planting porn seeds on a porn farm to harvest more pornsomebody gen this
>>102007920>>102007929bros you are killing it. seriously. bringing a tiny bit of light into my dark world. >>102007972NOTED
bonus if you guess the reference
>>102007956I did, but I'm guessing I need to do a pull from the folder because manager I did it via Manager.
>>102008029lol I did not prompt pain on the wall, that was just there
>>102008029NHK ni Youkoso?
>>1020080294chan banner, but also that other thing ^
So LoRAs with multiple character trigger terms aren't really working right now.
>>102008055you got it!
>>102008068Monkey looking fella
>>102008069You captioned your images wrong.
>>102008029>>102008071Room is too clean for a hikkikomori
>>102008086I'm not the only one experiencing this.
>>102008093They captioned the images wrong.
prompt: ifeelmyself
>>102008086no, it's simply harder to add multiple characters with loras, it's just how it is
>>102007769You don't want a base model that only has "good" images, you want as much variety as possible
>>102007972tried>>102008029>>102008044>pain>unpromptedflux gets it
Describing multiple characters would come from the T5 encoder, no? Do loras even train it?
>>102007802that clip_l doesn't work atm, use normal one
miku in a snow globe but also on a transparent cube:
>>102008069Works for me. I trained a LoRA on 2 subjects at once and it got it. Needs more steps though, and you need to have images of the subjects both together and individually.
>>102008140
>>102008134>triedi was imagining magazines not casettes but still good
>Hey guys! I can train a LoRA on 12gb of vram!>All I gotta do is brick my PC for 12 hours.
>>102008109is it "add multiple characters with loras" or "make a lora with multiple characters"these are very different things
>>102008213>he only has one pcAnon how poor are you? Don't you have a genning pc with 3GPUs and a separate gaming PC with 2 GPUs so you can gen while you train while you game?
>>102007628I tried extremely hard to do wow stuff but was quite bad. I never managed to generate a moonkin, even with all detailed description generated by LLM.
>>10200821312 hour nap
this gguf shit is fantastic. Wish it worked with ViT-L
>>102008213>>102008240
ideal amount of steps for good gens? 20 enough or no?
>>102008249is this base flux or with a lora? has more sovl than the usual slop it makes
>>10200833130 if you want consistency, 40 if you want consistency on text aswellhttps://reddit.com/r/StableDiffusion/comments/1er3wt7/if_you_want_a_good_compromise_between_quality_and/
>>10200833130-35 seems to be the sweet spot
Something got fucked up with flux loras on forge. It's crashing all over the place, but without loras it works no problem. Anyone else experiecning that?
>>102008366>forgeI see your problem. You are using experimental software.
>>10200833124-28. really depends on amount of ish in your gen. and your patience. text? add a few. lots of items? add a few steps.>>102008366>forge (sorry)
Hunyan is so much better than flux. Chinese engineers are #1.
So with Crystools in Comfy the meta data has the seed of the image I load, but how do I pull the seed out of it? I can get the prompt but I also need the seed
>>102008416>Chinese engineers are #1.Chicom industrial espionage is #1
does sd-scripts really not have a way to do validation images? how does one set it up to do that on an fixed interval with a variety of prompts? do they really expect us to train in the dark?
>>102008416It's a good model but flux definitely shits on it. Why do you think the entire community is using Flux instead of it?
>>102008462this
>>102008462it fucked itself from the start by having a retarded Chinese text encoder
>>102008477they fixed that?
>>102008332
>>102008485I honestly don't care, they should've launched speaking english from the start and now Flux is here and is better with full community support. Maybe they'll not fuck up the next time.
>flux having trouble with small stars of Davidif i prompted for a swastika i bet it would've worked just fine. how antisemitic of BFL>SPH2
>>102008486what prompt did you use?
>>102008497both work easy if you use high quality enough model plus t5 .. else it gets bonked
>>102008512photo of an extremely smug :3 catfaced chibi plushie anime girl with closed eyes and hands on hips, speech bubble comes out of her that reads "miku in a snow globe but also on a transparent cube", nuclear missile in the background, depth of field, blurry background photo of a chibi plushie anime girl with one eye closed and hand on hip open mouth pointing at the viewer, speech bubble comes out of her that reads "That is base flux.", big busty breasts oversized boobs in the background, depth of field, blurry background
>>102008417can't find a solution.https://github.com/receyuki/comfyui-prompt-reader-nodetry this. why do you want the seed tho?>>102008486really good texture.>>102008497lollipop fry
>>102008555>https://github.com/receyuki/comfyui-prompt-reader-nodeThanks I will try it.
>>102008555Oh I need the seed to upscale, I don't want to upscale from a random seed, I want to use the exact one, to avoid any changes to the image I am upscaling. There's a tendency for upscales to add and remove details.
>>102008497
>>102008637a daring synthesis (assuming it's the buddhism symbol)
>>102008637should be lgbt colors. or ukranian flag lol
>>102008555>lollipop fry?>>102008559>amateur tv screenshotsunironically a genius use of flux. more creative than surveillance footage
>>102008398>>102008411oh well
>>102008643You can just ask for a nazi swastika and it will give it to you
>>102008672I haven't had durst styli in ages.
>>102008672I saw someone use cctv as a prompt and was amazed, the model can do cool shit
>>102008684you can also just ask for "Star of David" and it will give it to you
>>102008701it keeps surprising me, I'm slowly snapping out of my WD tag mode and really pushing it, describing exactly what I want because odds are it's in there somewhere, even though it gets pushed tot he back by all the other far more common things
>>102008637i wonder if it can do the Raelism symbol
>>102008715>Raelism
now this is podracing
>>102008619no no do not use the same seed for a resample/upscale. the moment you resize the image the seed becomes useless. worst case you burn the image. forget about the seed. trust me.
https://xcancel.com/MrDavids1/status/1825208573543981111>I've been secretly testing Mystic which I think is even better than FLUX with AI image generation.Will it be local though? if not gtfo lol
>>102008672
>>102008559Lul
what's a good sd model for fixing anime hands specifically? I got a nice looking depth map from hand refiner but that's as far as I'm getting
>>102008770need to be more specific. sd15? sdxl? pony? anus1111? forge? comfy? mesh graphormer? its always a hustle to dial it in
>>102008730I trust my eyes.
>>102008815sd15, comfy, mesh graphormer. depth map preview looks perfect, but after feeding it to a controlnet and inpainting I just get different blobs. I suspect my model just can't do hands well regardless of controlnets
https://civitai.com/images/25424379MJ lora got updated, that's cool, I hate the generic anime style flux always outputs
>>102008804
>>102008854sd15 can into hands but you have to wrestle it a lot. It's extremely outdated at this point.
Oh god I fucked up I am genning a 786,484 X 786,484 image
>>102008726the symbol got renewed interest when Kanye posted it on twitter>>102008878the half-burger with fries on top looks appealing. surprising no one has ever sold something like that
>>102008890Seems like one of those times where the NVIDIA fallback feature might actually fuck you over and crash your PC
>>102008890Grab your fire extinguisher.
is the beta scheduler better than normal?
Best use of ai for me is creating wallpaper
>>102008879yeah I just jumped straight into my old setup from over a year ago. not sure if I can handle anything better with 4gb vram
>>102008904>the symbol got renewed interest when Kanye posted it on twitterthe shit storm was so huge tho that I don't think it found its way in data set, when I prompted it it just made strange random symbols
>>102008908All I know is I saw it on time and killed comfy.Also it was much larger because I fucked up elsewhere. It was an upscale doing 16x16 tiles too. So it was really 256 786,484 x 786,484 images
>>102008941And creating noods of friends girlfriend
>>102008854lets see..- controlnet strength somewhere between 0.4 and 0.7-step size around 12, scales with denoise. (lower denoise=less steps)-and then the all important value, the denoise, needs careful adjustment.
>>102008904I didnt ask
just deleted all of my models + gens lads
>>102008943You could really try and push Flux and use Q2 unet with Q3 text encoder. I expect it to look very shit at that point though
>>102008995if that's the average quality of your gens not much was lost
Morning
>>102009009hello stranger
can someone point me to a workflow for flux which includes the option for a lora, I just trained one on civitai but im so used to a111
>>102009014here have mine: https://files.catbox.moe/28y5hz.pngit aint pretty, but you got options for three loras, dynamic thresholding and on the right side even sd ultimate upscale
>>102009014default workspace, add a lora loader (loraloadermodelonly)checkpoint -> lora loader -> ksampler, and you are good to go.
>>102007920>how do generate boob,
>>102008957thanks gonna play around with this, a few tricks here I wasn't doing, and better starting values
so does flux "converge" at 20 steps? is 30/40 any better or diminishing returns?
>>102009146it was trained on 1000 steps so I'm doing 1000 steps.
>>102008941Looks nice anon, reminds of the default KDE Plasma wallpapers.
>>102009146not on all samplers, not on all topics, I stay with 22-25 most of the time, but 20 is good enough mostly
there we go, perfect ad.
>>102009180kek
6.3gb vram used. Q3_s t5 and Q2 model.I think this is the absolute minimum for Flux.
>>102009259have you tried text with that combo yet?
>>102009280
>>102009310nice. there's hope for 8GB vram users yet! thanks for posting
>>102009332Could start a gen on flux and upscale it with SD to clean it up so you get the good prompt understanding and some quality.
well that was 4 hours wasted
>>102009373Are you trying to untrain the butt chin?
>>102009383no that's supposed to be holly willoughby
there we go, I wanted someone holding a book. >>102009286
>>102009390how do you fuck up so hard
>>102009413I believed the lie
>>102009390Oh right, yeah definitely wasted 4 hours, but, you now have a LoRA you can use to produce a consistent looking woman that you can scam men with.
>>102009392wait, better:
why does flux image gen take 10x longer with a lora, is that normal?
>>102009373you definitely fucked something up, cause its super easy to train flux on faces
>>102009442if you got buzz
>>102009436You ran out of vram and the overflow is calc'd in RAM instead (slow). Use smaller loras
>>102009449It costs less to buy buzz than the power it costs to train one yourself. That's how they get you.
Really interesting, AutomaticCFG seems to be also a viable option for CFG > 1 (With Tonemap and DynamicThresholding)https://imgsli.com/Mjg5ODE4/0/3https://github.com/Extraltodeus/ComfyUI-AutomaticCFG
>>102009473Oops I forgot the prompt:>Hatsune Miku and Sailor Moon having a cooking competition, with Miku making futuristic dishes and Sailor Moon making traditional Japanese food, pixel art style
>>102009449i do it local on my 4080, takes me about 1.5 hours for a lora that recreates the subjects likeness pretty well ~500-700 steps
>>102009524I don't know who that is supposed to be
>>102001109You're perfecting your technique holy shit.
>>102009524will you please share lora settings?
>>102009259im on q5_1 with the same vram, using forgealmost 3mins for this generation, expanding it takes 15min
>>102009524How does it feel to pay $1500 just to generate mindless slop without any artistic value whatsoever?
yoji shinkawa style (mgs) lora, this is cool:miku on the mgs3 title screen
i trained a lora on civitai, I gave it about 25 portrait photos of this woman, and for some reason all the lora generates is dog pictures, wtf is going on, in the preview images one of them was a dog with the head of the woman I originally gave it but besides that it's just churning out nonsense
>>102009490Hardly a technical prompt. You are an idiot if you think this has some merit in terms of finding out about techniques.>>102009656>loraYou don't need lora for that and besides, this is nothing like Shinkawa anyway, retard.
>>102009668no u
>>102009652a 4080 isn't $1500
>>102009668you must be some special kind of retard to not notice the difference between cfg 1 and cfg 6
Smell that fresh loaf of bread straight from the oven...>>102009692>>102009692>>102009692
>>102009652>artistic valuenigger our GPUs are drawing this shit. Get the fuck out of here.
>>102009652feelsgoodman
>>102009652Good.
>>102009594>I don't know who that is supposed to beits just a random egrill, so that doesn't surprise me>>102009603i just use the recommended settings on kohya's github with the <12gb optimization launch arguments. I've done 10-15 epochs with 20-30 image datasets, captioned with joycaption and prepended with a trigger word.Halfway through baking the loras start to work well, but obviously look better if you give it more time. The 10 epoch 20 images took me like 2.5 hours and worked pretty damn well. 15 epochs 30 images took me overnight ~8 hours>>102009652I built this pc at the end of 2023, upgrading from an i7-4790k + 970. I didn't fomo into a new GPU for AI, it was just time for a new computer.
>>102007828you can train 1024*1024 on 12gb, so
>>102009876neat. i downloaded a config for kohya and it was set to 512x512.
>>102007980>halo3 scirtless>master chief teleports behind the skirt wearer with an unforgiving look in his eye>his left hand closing inuh oh...
AdamW8bit or Adafactor?
>>102009682catbox of this one pleasei love the artstyle, is it a specific lora?
>>102010223thanks I can't stop laughing
>>102010630nvm i didnt see the other post u replied to, gonna give this lora a try
>>102009876But wasn't it yesterday or something like that where several anons here were making fun of some anon that said you could train on 12gb vram? what happened?
>>102011220It was just d*bo stirring shit.
