Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101935309>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>GPU performancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>Maintain thread qualityhttps://rentry.org/debo>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>>101937500Was about to mention, this, I can't daisy chain two LoRAs together with gguf right now. Which sucks. Luckily I'm not a total vramlet.
Im just like gonna call it slop haha
L O R A S O R A S
flux was worth it, solely for the Pepe lora. The rest is a bonus. And this is basically the first week of loras, most aren't even training yet.
So I'm interested in captioning my datasets now for LoRA training to see if it can improve the outputs. What captioner are we all using right now?
>>101937515>can't use one lora without tonemap and therefore without cfg > 1>with cfg = 1, can only load one lora and not moreyay...
>>101937590for SFW, GPT4V, for NFSW, that joycaption thing
>>101937606>joycaptionWere there weights available?
>>101937590Joy captioner DL from huggingface and replace with model another llama model to run locally/uncensored
>>101937628>replace the model I'm going to bed it's 3am
>>101937620yes, just clone https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alphait will download the weights for you (the llama weights require you going to the repo and requesting permission but there are mirrors on hugging face and finetunes work as well from what I've seen other anons do)
>>101937574slightly better: forgot the cartoon promptremember, if you can generate controversial stuff then the basic stuff is easy as fuck.
>>101937495From a technical point of view, how does Flux differ from SDXL?As in, if someone was trying to adapt something that supports SDXL to also support Flux, what would the key points of difference be?Two CLIP encoders (is the second one optional)? What other magic is there?
>>101937640lmao that one is good not gonna lie
>>101937641it's a diffusion transformer architecture instead of a UNet, does you "something" care about the architecture?
Pepe is wearing a business suit and is at a political rally. He is standing at a podium. A big rectangular sign that says "PEPE" is behind him. Pepe is saying "feels flux, man" in a white speech bubble.
Quite a few flux loras are beign trained on AI images, WTF is wrong with those people... It just makes everything look more AI.e.g I saw a lora and I was like why does this look like a Pony image...well they trained it on Pony images.
>>101937673Do one with a bullet grazing his head.
>>101937638I hope they add this to Taggui, I love using that program.
>>101937690those retards find now normal to do some AI inbreeding, I fucking despise that shit
>>101937691
>>101937732slightly better:
>>101937661Ah right, that's probably going in the too-hard basket for me then sorry.Thought maybe I could hack support into StableDiffusion.cpp.
>>101937641>From a technical point of view, how does Flux differ from SDXL?better transformer architecture (DiT vs unet)Way bigger (12b vs 3.5b)
what a time to be alive, huh?
Is there a way to drag an output that is already linked so that I can choose a new output instead of a new input? Haven't found a modifier that changes it from the default.
>>101937732>>101937748fire the whole bulletthat's 33% more bullet
>Hatsune miku, her speech bubble says: "I'm loving it!", 50's comic book styleChoose your best unslop combinaison (X = cfg | Y = GuidanceNegative), made possible with Tonemaphttps://files.catbox.moe/a8qovd.png
>>101937764>>101937748>>101937732KEK, nice.
>>101937813No, kill yourself
It's over
Anyone knows how to change the font size of the prompt text? I edited the user.css withelement.style {--comfy-textarea-font-size: 20px;}But it gets overwritten by html when opening comfyui
>>101937893It wouldn't be ideal, but could you overwrite it in inspect elements?
>may I take your order?
>>101937900I did but it goes away after I close the tab
>>101937909use Stylus
>>101937905
>>101937893Isn't it an option in the settings? Click the gear and look for "Textarea widget font size"
>>101937934Holy fuck it works, tried for like 2 hours to change it
the colonel in the background is the most impressive part.
>>101937944I wonder if by virtue of most KFC marketing having the letters KFC in them, they came out better tagged in the dataset.
>>101937934Thanks, now I need to find a font for it
>>101937944brands come out pretty good in addition to text too
is there an example repo of someone setting up a flux LoRA with a dataset so I can just copy it without thinking too hard?
More steps in schnell adds more details
>>101937515>>101937591I fixed it kek. Who needs sleep.
>>101937996looks like a Bratz doll
>>101937905>anime 2bCan it do a normal in-game one?
>>101938020this lora is trained in that style but if you gave it an artstyle prompt I guess it would work, not sure still testing.
why is my VRAM usage higher today and why does moving the VAE to the CPU give me OOM now
>>101937997It works!! No errors and no OOM anymore, what a fucking legend!! https://www.youtube.com/watch?v=VtAzlUu4b3Y
>>101937996Truly groundbreaking, Anon.
RAYGUN COULD NEVER>even 2b is better than the aussie
>>101938045but can 2B hop like a kangaroo
>>101938045this time but with more soul and the logo
>hit the new "queue" button>comfy freezes and lags for a couple secondsGuess it needs gpu acceleration on to not shit and piss itself
>>101937997No more errors anymore, for those having OOM, remove --highvram and remove the force MODELDevice thing too
>>101938094But gpu accel directly affects your generation speed, why would anyone enable it
gn everyone
>>101937638so basically we need a bit of a wrapper to come up with a dataset to avoid duplicating work
>>101938045nice
>>101938181we can use AI to make the routine a 10/10
I do Schnell 1024 res gens in about 18 s on my 3060dunno if that's slow or notRAWRRR
>>101937813
>>101938191That's normal for now
https://civitai.com/models/644109a new era of pepes begins>>101938211very nice!
pepe is also a fan of miku gens in flux
What scaler or node can do this?
>>101937997Works perfectly with various loras/cfg adjustments and whatnot, thanks. Custom loaders work just fine too so no need to stick to the regular lora loader.
So GUUG Q8 works better with loras than Nf4 right.
Flux 1 dev with Realism Lora + Magnific + Luminar + Lightroom
>>101938282it gave miku pepe eyes kek
>>101938290GGUF*
>>101938291a lot of work for a noisy analog
>>101938290it should, Q8 is way better than nf4 on its own
>>101938308Sounds good, downloading now baby
why the fuck is ComfyUI using more VRAM today, it's starting to hit the shared VRAM when it has been working fine for a weekeven changed the DE to the iGPU, the VRAM use on the GPU is literally 0
Can you inpaint faces in Forge yet for Flux?
>>101938315Did you update it? I noticed the last update did that
what a time to be alive>no there will be no nudes>no loras are impossible>no you need an a100 to run it>no it has safety controls on it
>>101938315Are you using quants? Update the node and comfyui. idk which fixed the issue
>>101938262eh this isnt magic. just looked at their images. upscale by model (pick a decent ESRGAN model) > downscale to a reasonable size > resample with a low ish denoise at around 16 steps > upscale again to final size - with flux you get actually better results lol. many ways to skin that cat bro
>>101938331I wasn't but tried it just now to see how well it worksusing Q8 the VRAM use is lower so no swapping but it is two times slower than FP8, from 2.4s/it to 4.86s/itis that normal?
>>101938348>using Q8 the VRAM use is lower so no swapping but it is two times slower than FP8, from 2.4s/it to 4.86s/it>is that normal?usually when you get something 2 times slower that's because you went from CFG = 1 to CFG > 1, the speed between Q8 and fp8 are supposed to be close
>>101938362I don't use CFG at all so it's not that, I just changed from the normal loader to the GGUF loader
>>101938371maybe it's overflowing your VRAM capacity and that's why it's slow, how many vram do you have?
>>101938390I have 16GB>>101938348>using Q8 the VRAM use is lower so no swapping
>>101938020To answer my own question - kinda works but it definitely wasn't trained on the game renders. Looks more like a gta san andreas mod kek. I just put "a 3d screenshot" in the positive and "anime, 2d" in the negative
>>101938362is the gen process pretty much like this:comfy sampler -> gguf dequant -> flux model script using NN Linear?thanks for the great work
>>101938404>I have 16GBdo you see if it overflows your vram capacity on task manager? Piercel is the picture of Q8 vram usage during inference on my 3090
>>101938430I literally said it doesn't in the comment...
Ingore the audience
>>101938406And this is with "a real photo"
>>101938439welp, sucks to be you I guess
>>101938334Some dev ones look similar but takes forever to generate compared to schnell
>>101938457you can prevent that by adding "anime" in the negative prompt
>>101937996Take the 10 minute per gen pill. Go dev. You will learn to discern if it's worth it by looking at the preview of the first few iterations, and you will put in a lot more work into your gens.
isnt technology cool, im a comp sci guy, not an art student.
>>101938457Not bad.
>>101938470Flux is worth it just for the coherence of backgrounds
>>101938478flux has a way of making photos/depth of field type prompts feel authentic, i'm not sure exactly how to describe it, but it works. It uses noise in a better way than SDXL, I guess?
Why doesn't image2image work with flux in general? Or is it just with schnell?Image gets really overcooked or just broken
that explains it https://github.com/city96/ComfyUI-GGUF/issues/33#issuecomment-2294821171I read that Forge also slows down with loras, is that right?
>>101938468Yeah, that's the first thing I did.>>101938510Everything always slows down with loras, even 1.5 and sdxl.>>101938497Dev needs a pretty high denoise value for significant changes, I'm talking 0.8+. No idea about shnell.
>>101938497I2I should work with every model
>>101938524>Everything always slows down with loras, even 1.5 and sdxl.never by doubling gen times as it is happening now
>>101938348Something similar with me not double but more, went from 3.9s/it to 5.5s/it. From 1min 20 seconds to 1min 50 seconds.Gonna restart PC to see if it helps
>>101938191That's normal. My schnells take 1-3 minutes on my AMD 5700.>>101938495What the fuck is going on in that back room?
>>101938460I scale them to 1.5x (of the original) size after the "upscale by model" node before resampling it, its reasonably fast. then, as a final step, in a seperate workflow, I just upscale with a clean ESRGAN model again and downscale to the final size. tried various DAT upscalers, waste of time in most cases. really hungry mofos. BRO
whats a good workflow for Lora stuff?
>>101938536>never by doubling gen times as it is happening nowTheorically it should double the gen time because what CFG does is creating 2 images instead of just 1, so twice the work
>>101938495and for something a little different, art nouveau lora but with miku:
>>101938549motherfucker I'm talking about loras what the fuck are you on about
I got a great deal for 4070 ti super 16gb model, anyone using it for image gen?
>>101938542I see a booger up her nose. Realistic.
>>101938555>Seething this hardTake some pills retard, you need to control your anger issues
>>101938495requesting catbox pls
>>101938574you read "doubling gen times" and your monkey brain thinks "CFG" disregarding the context of the conversation because you make logical connections at the level of GPT-2
>>101938566it keeps happening on this prompt.
>>101938556Using one right now for Flux. With ComfyUI's low VRAM mode you can run the full FP16 model.
>>101938588Like I said, take some pills, if you're gonna scream like a monkey everytime someone missed something in a conversation you'll have some serious health issues
>>101938582https://files.catbox.moe/tgka91.png
>>101938548Just put a lora node in.
>>101938603So you admit you're retarded.
>>101938556BRO EVEN A 3090 OUTDATED NOW (t.3090 user). seriously don't waste money on anything below a 4090 now
>>101938588That dude spends his entire day posting his CFG hacks to here and reddit, any chance he can he'll shill his shit. His images always look deepfried and uber noisy because of it too.
>>101938620So you admit that you have the same self-control as a unhinged gorrila?
>>101938536About 25% speed decrease or so for me. Running Q4_0 on RTX 2080
I tried the pepe lora with the nf4 model. I got kermit instead. lmao
Is it me or her legs are looking a little thick.
>>101938636Damn, nice
>>101938556>>101938623Everything is "outdated" when it comes to this. I'm going to get a 24GB used GPU myself as soon as I find a good deal. Otherwise you will be forever waiting for the next thing and regretting spending money because this scene moves so fast. Just make gradual upgrades and don't get to hung up on it.
>>101938595Hey that's cool. How are the gen times?>>101938623I got a deal for 400€ vs 4090 almost 2k€ where I live
>>101938556I have a 4060ti 16GB and 64gb ddr5 and it does the default fluxd 1024x1024 in 49s, with comfyui on linux so you should be in for some good times.>>101938611Bless you anon, doesnt work for me, using comfui, but bless you anyway.
>>101938623That attachment on the tip of the gun is for sticking it somewhere.
I have no idea how I got outputs this varied from:"Alex Garan artstyle, Page from American college yearbook, , 1980s in color"
>>101938684means it has fuck all idea what "Alex Garan artstyle" is
>>101938669Dale duro hermano. I'd go for more memory if I could, but that's a hell of a deal.
>>101938670it's not complicated, it's just the emma lora then after the instance prompt "in a coffee shop, she is wearing a white tshirt that says..."
Same prompt.
>>101938669For that pic, which was 25 steps with Euler, it was around 45 seconds.
>>101938682fact checker
so in comfy, how do you connect the gguf -> lora -> nodes to text input? the output after lora is model, what does it link into?what's the basic workflow for gguf/lora?
>>101938669>I got a deal for 400€fucking how, my 4060 Ti 16GB was 450the 4070 ti super was ~850 at a minimum last I checked
Say what you will about the comfy UI design but I am absolutely loving these node timers.
>>101938720The lora alters the model. It just makes it react to new embeddings. You don't need to connect the lora to the text/embeddings input. Think of it as a filter for the model so that it understands your toddler foot fetish.
>>101938720*cause the initial workflow has clip and vae I believe but the q4/q8 loader doesnt
>>101938720depending on the LoRA loader you use, there should be a yellow clip thing you can pull out into the text input.Depending on the LoRA though, it's probably not necessary.
>>101938495If you can't beat them, join them
https://civitai.com/models/523485?modelVersionId=732778Finally, flux is perfect now
>>101938735>>101938746thanks
>>101938720That said, there are several lora loader nodes. Some let you plug the image model and the text encoder (clip). I dunno.>>101938755My fetish.
https://civitai.com/models/653149/javelin-82SHUT IT DOWN
you know not even a few months ago I would have said open source DALL-E 3 was pretty far away, and now flux is basically uncensored open source dalle. pony/sdxl was already good, this is a step above that.openAI/Microsoft might be weight hoarding assholes but open source always wins.
>>101938614Sir pls I need the workflow
>>101938755>dataset consists of 28 screencaps (1920×1024)I hate these people so fucking much. Same as the ones finetuning text models with arbitrary prompt formats.USE THE SAME FORMAT USED ON THE ORIGINAL MODEL FOR FUCK'S SAKE
>>101938665you are right, yes. that's why I got this 3090 last year. I don't regret it. still, feels too weak sauce for flux.>>101938669my advice: spend a bit more and get a 3090. 500? 600? you can do it.>>101938725it is going forward.>>101938766kiki best girl
>>101938779Lmaoo, how the fuck civitai is still alive after hosting so many controvertial models
>>101938786Flux was trained on a range of megapixels and ratios thowhat you should be saying is USE A VARIED SIZE AND RATIO DATA SET FOR FUCK'S SAKE
>>101938725I like that too. I also like seeing the green outlines so you can visually see the program at work, it's like seeing data go from your ssd to your cpu and memory or whatever.
>>101938779What's the trigger phrase for this?
>>101938786>USE THE SAME FORMAT USED ON THE ORIGINAL MODEL FOR FUCK'S SAKEThat would ruin the whole thing, the ghibli screenshots aren't 1:1, flux must know how it really look like
>>101938725how do you get that anon? that looks interesting>>101938803>What's the trigger phrase for this?kike
>>101938812Just crop the images judiciously. It's not about genning "a screenshot of a ghibli movie", but genning "kiki getting all her holes filled by a gang of magical brooms".
>>101938801It always had the outlines though
>>101938812but it knows cinematic wide shots alreadybesides the full shots you could also train on crops of the characters (keeping them withing the size/ratio Flux knows of course)would bulk up the data set and avoid the chance of it degrading with gens that aren't wide
>>101938843>Just crop the images judiciously.that's fucking retarded, flux is good at a lot of resolutions, so why not train it further with the real deal instead of forcing it with some 1:1 nonsense
>>101938843BRO.
>put in the Load Lora node into the workflow>press queue prompt>nothing happensI dont get it
>>101938785https://files.catbox.moe/qjmvkr.png>>101938825Update comfy and enable it in the settings
>>101937495How's the Flux.1 q4_0 compare to q4_1?
>>101938698ah i didn't see it, got it now (was being a filter-tard) ty
>>101938843Most training programs auto resize and crop pictures anyways to fit the SDXL resolutions, no need to do it manually
>>101938859Q4_0 is always inferior to Q4_1, in the LLM space, Q4_1 has been created to improve from Q4_0 by adding some more high precisions into the weights, so it's a bit bigger but better too
>>101938856You need to say the magic words.>>101938851If you made the LoRA, know that there's a reason you had to inpaint that fucking third leg out of that image. If Flux is giving your anatomical errors in 10 images you made as samples for your shitty work, there's something you did wrong and you know it.
>simple guidance>adaptive guidance>dynamic threshold >guidance threshold>flux sampler>flux shift>two CLIP nodes for positive/negative with individual guidance settings>skimmed cfg>tonemap>CFG Guider>AutoCFG>PrepNegGuiderI am lost.
>>101938885Worse yet, you are going to get randomly cropped faces and bodies that way.Just put some effort into your work if you're going to publish for other people ffs. That's all I'm saying.
>>101938889Don't bother with that one anon's setup, it's hacky and breaks whenever a new update comes out. Just use the basic workflow for now until better supported CFG options come out
>>101938889not enough nodes sir
>>101938889When did you start learning on imagegen anon? You should try easy first and then learn each concept slowly but surely
>>101938889>don't need it>don't need it>don't need it>don't need it>don't need it>doesn't make a huge difference>don't need it>don't need it>don't need it>don't need it>don't need it>don't need it
This is bad, right?
>>101938919you don't need dynamic threshold? seriously? that shit is great at unslopping anime images
>>101938921you're treading on dangerous ground
>>101938901Considering everyone already uses those tools and things turn out perfectly fine, I'm gonna have to go with you're just sperging out for no good reason. You can mess around with the tools used for bucketing/downscaling/cropping, you'd be hard pressed to throw something at it that will truly get fucked by the process.
>>101938933Not intentional, but I can think of the implications of this from a LoRA.
>>101938919Why are you saying you don't need Adaptive Guidance? You improve the speed of your gen with that, smells like some serious skill issues if you ask me.
>>101938921HALF GIRL HALF PLASTIC BRO. & she already hit the wall there
>>101938955>You improve the speed of your gen with thatonly if you're using CFG, you don't need it
>>101938945>things turn out perfectly fineYou have real difficulty finding badly made LoRA's on Civitai lmao
>>101938971nta, but CFG is good to remove the blur of photos flux is producing, I hate that blurry shit, why does flux have such a bias towards that
>>101938857wat do?
>>101938932It's good for anime, but for anything else it adds even more slop. The guy who shills it uses the same "woman on a street" picture to test his settings because for 98% of anything you'd want to prompt photo-wise it's trash
>>101938983delete the nf4 loader, I have it there for no reason
>>101938983It's over.
>>101938921did you accidentally put megamind in the prompt?
>>101938990>It's good for anime, but for anything else it adds even more slop.So it's useful and you need it after all, unless you're pretending people aren't into anime or something?
>>101938979Civit slop will be Civit slop no matter what tools you give them unfortunately kek
>>101939003>It's good for anime>If you use it for anything else it's trash>So you need it then?Not great at reading between the lines their kid are ya?
>>101938919>>don't need it>>101939032>>It's good for animechoose one anon
>>101938755What the fuck are these birds
>>101939038>Don't prompt for anime>I don't need itNo :D
>>101939042That's something you haven't specified on your first post, weird huh? :^)
>>101939038>me>not meyou people really can't tell when you're talking to different people?
>>101939040cummy birds
>>101938918When CompVis released 1.2.Still way too many new nodes without very much documentation.
>>101939040birds aren't real
>>101939059Considering the guy only seems to talk when he can shove his workflow into a convo it doesn't surprise me that he doesn't know who he's talking to
>>101939040That's the cum of everyone who ever came to kiki's doujinshi, impregnating the imaginal space from which these images emerge.
>>101938706Ty>>101938724I order it using friends business that gets good offers every now and then>>1019387913090 way too expensive where I live
>>101939059Bullshit I can see your name right there "Anonymous". You've spammed this thread.
>write running behind a bus>the bus is always behind himI thought this thing understood natural language. This is basically the same issue SD had.
>>101939065>Still way too many new nodes without very much documentation.The only documentation we have so far are the tutorials on reddit, we definitely need a rentry or something
>>101939114clearly he's running behind the bus that is behind the cameratry describing the back of the bus
>>101939114This being a flux gen, you can clearly see it's trained to produce the kind of images that wow normies and journos. 1girl instagram simulacra. The moment you try to describe something that isn't that, it starts introducing weird shit like a shit model. Look at that foot. And what's going on on the left? It also hasn't decided if the camera is inside or outside a bus.Flux is overtrained on specific types of images to make it look good on social media.Prove me wrong (seriously, post non-typical situations)
>>101939114try to go for some boomer prompts, chatgpt or claude can help you on making your prompt more descriptive
>>101939114what about running after a bus
>>101939159Damn, your life must be harsh at that moment, you're forcing yourself to be on /ldg/ and see people only talk about flux when you could've gone back to /sdg/ and enjoy talking about your favorite SD models, why are you doing that to yourself anon? genuine question
>>101939114https://www.youtube.com/watch?v=MjbUnn32_zU
>>101939159>when you try to describe situations it doesn't understand, it doesn't understand themWoah
>>101939040the humans on the ground look horrific aswell, must be a terrible lora to make flux shit itself like that
>>101939159That's why I upscale and inpaint with SDXL. Flux is vanilla
>>101939169I know, I'm just trying out things. That prompt was "a jewish man running behind a bus".This one is "a bus behind a jewish man running". Just interesting to see how this thing thinks.>>101939191Probably more idiomatic.
>>101939191>a jewish man running after a busThis prompt breaks flux apparently.
>>101939040The more LoRAs you mix the fuckier things seem to get
>>101939260that's why we need a finetune that adds more concept to flux, doing some LoraMAXXing has some serious limits, what if you want to go for 2b + Pepe on a ghibli style, that's 3 Loras flux has to handle, it can't work
>>101939260Lycoris doesn't work with flux?
>>101939292Asking the wrong guy here. I think most people are still kind of stumbling over the finer points of LoRAs for flux.
Here our tormented Jewish man is chasing the bus in the opposite direction
>>101939321that's a muslim
>>101939318IA3 should also be tried.
simplest way to train a Flux lora on a 4090? want to try and improve its accuracy in generating a specfic outfit
>>101939328They're all semites, what difference does it make?In this one I used "chasing a bus", and he didn't even bother running.We need a LoRA for this it looks like.
>>101939345https://github.com/ostris/ai-toolkitWorks on windows. Just follow the instructions and you'll be cooking your LoRA within 20 minutes.
>>101939363>Just follow the instructions and you'll be cooking your LoRA within 20 minutes.So that's why there's a spam of Lora on civitai recently...
>>101939350the model is poorly tagged, embeddings would save the day here, yet that entire tech is being ignored atm.
>>101939373They are all very good quality because of the reasons this anon stated >>101938945You can train LoRA's using any resolution, any number of images, and have them automatically cropped by a script. And it will come out amazing and ready for Early Access monetization!
>>101939363>https://github.com/ostris/ai-toolkitEvery time I have tried this I get an error. Gonna delete it all and start again, see if it still fails.
>>101939373>So that's why there's a spam of Lora on civitai recently.No. Even that's too difficult for the average smooth brained jeet on civit. Civit recently introduced in site LoRA training in exchange for "buzz" (whatever the fuck that is)
>>101939394>You can train LoRA's using any resolution, any number of images, and have them automatically cropped by a script.like it will adapt to each resolution each image have? that's an excellent feature if you ask me
>>101939400>Civit recently introduced in site LoRA training in exchange for "buzz" (whatever the fuck that is)Imagine giving your data + captions to civitai for free, I don't want to sound rude but you have to be pretty retarded to do something like that
This is the best attempt I could muster. Gonna try Pro, see if it makes a difference
>>101939426>for freeActually you have to pay for this too.
>>101939428
>>101939426Really depends on the data and your hardware? Me? Never in my life would I fork anything over to civit. But the random jeet might want to upload his face five times to make another linkedin scam and that doesn't really bother me.
>>101939441I won't be surprised if one day Civitai will make their own giant finetune of flux out of people's data and make their own API from that
>>101939428ImageFX is better trained in many ways
>>101939483Can it make a wet woman?
Not getting spellcheck in the prompt box in ComfyUI for some reason nowI'm ESL, I need it.
>>101939363If I wanted to train my own LoRA from nothing but a bunch of images, what's the best way to batch caption them? JoyCaption? (does it even run locally?)They all have the same [item] in them if that matters, not even sure JoyCaption would recognise it, and I want to ensure it captures every little detail.Never done this before so any advice would be appreciated.
>>101939523yes but you have to tard wrangle
>>101939564>does it even run locally?yes>what's the best wayby hand>They all have the same [item] in them if that mattersYou could state that in the prompt
>>101939564>If I wanted to train my own LoRA from nothing but a bunch of images, what's the best way to batch caption them? JoyCaption? (does it even run locally?)JoyCaption do run locally, I still think GPT4V is the best captionner model, I would use GPT4V for SFW and JoyCaption for NFSW desu
>>101939564Well, if it's nice enough that JoyCaption won't even recognize it, then you can run it through it and then edit the outputs manually to fix any mistakes or do it raw.It's also worth considering if you even need captions at all. Are you training a brand new concept or just trying to enhance something that's already in the model or force a particular style? If that's the case, captioning might be completely optional.
>>101939428on a streetlooking at the tail pipe on the back of a buslooking at the back of a man with grey hair sprinting after the bus
10th try and I'm still not happyit wasn't that complicated of a prompt
>>101939604LOOKS LIKE A JEWISH LADYBOY BRO.
>>101939604>it wasn't that complicated of a promptStick it in GPT and ask it to rewrite it to be more verbose?
>>101939397Still failing. The file is there so I don't really know what the issue is.
luck of the draw I guess
>>101939658I do. Where is your config file in the config folder?
>>101939670catbox? want to slap a realism lora on top for better skin
>>101939670it doesn't look like Emma at all
>>101939428this seems harder than i thought. need to ask claude
>>101939675In that folder.
>>101939670
>>101939682>>101939601
>>101939681Emma? isn't that Taylor Swift?
>>101939687Can I see? Something isn't adding up.
>>101939713Oh fuck you're right, kek
I wonder why Lykoris never managed to be the replacement of Lora, it's in theory a superior method
FLUX doesn't work on A111 webui, but will it ever? Is that even possible?I like A1111...
>>101939721I see the previous Emma image now so I can understand the confusion.
>>101939714
>>101939697Using dev + schnell combined for 4 - 8 steps, seems betterhttps://huggingface.co/drbaph/FLUX.1-schnell-dev-merged-fp8-4step
>>101939740why won't you jump to the forge ship, it's A1111 but faster and with all the recent updates
>>101939740just use Forge
>>101939574>yesHow exactly? Last I checked it was just a HF Spaces frontend to a model, not sure what tool I'd need to run it locally. >>101939578>GPT4VEh, that's paid so sadly I'll have to stick to JC. Hoping it'll suffice.>>101939589Just tested and JC recognises it, fortunately.I'm actually just trying to enhance the details of a certain historical character that the model is already capable of generating, but not 100% accurately. For example, an emblem on the character's cap always looks wrong or a badge on his suit sometimes looks off, or even his general appearance since I sometimes have to generate 5-10 times before a decent result. I'm guessing that would require captioning.
>>101939752until it gets abandonedcan you pin T5 to the cpu in Forget yet?
>>101939745Okay, I see the problem.your config file is named naif.yml and you're trying to get it to run a file called lora.yaml which doesn't exist.try this: python run.py config/naif.yml
python run.py config/naif.yml
>>101939765>can you pin T5 to the cpu in Forget yet?Nope, that's why I'm stuck with comfyUi, this shit has more important features
>>101939761>How exactly?huggingface spaces can be cloned, they are git reposthe app downloads the models for you (the llama model from the official meta repo requires asking for permission first but there are mirrors you can use just by changing one line in the code)
>>101939778wait fuck. I see in your screenshot you've changed the name.
>>101939699>tail pipe on the back of a busnice. adding "tail end of a city bus that's pulling away from its stop" did the job
>>101939778no I renamed it. check the terminal again, it's calling the correct file.
gentlemen
>>101939752>>101939754I haven't done this in a hot minute, didn't realize there was a better fork, thanks anons.
>>101939797I even tried calling the full path
>>101939802they thought they could stump T5 when it was them who were promptlets, although "back of the bus" should be pretty clear
>>101939114My one is running head on into the bus...
>>101939578What about stuff like llava?
>>101939805Okay. One more thing though. In the command prompt you are writing "niaf" when the actual file name is naif
>>101939813it's impressive how well flux is learning new concept, and this is just some loras we got there, imagine a serious finetune on this, I hope someone is up to the task, is pony-dev hinting he'll be working on that model?
>>101939840someone will do it, it's a capable model
>>101939840>is pony-dev hinting he'll be working on that model?he can't monetize FluxDev so no
>>101939859he can on schnell, that model fucking sucks though
>>101939840>is pony-dev hinting he'll be working on that model?Nope, he's been adamant about not wanting to do.
>>101939834Oh holy fuck I am a retard. You're right.Now I am getting an error about the folder path of the files but that can be fixed, probably doesn't like the spaces.
>>101939884No problem. I once spent like two days trying to figure out an error before I realized I spelled the world university as unversity.
I can't get inpainting to work with GGUF Q8 in forge, hmm.
>>101939897And word as world.
The guy laughing and pointing is a nice touch if I do say so myself.
>>101939897Ifs a feeble mind that kant think of moar then won way to spel a wurd.
>>101939928I know. It's my weakness.
this retard https://civitai.com/models/647663/porsche-911-gts-2024-flux>2.5GB lora>the LoRa is this big because it is trained with really high res images think 4k and with a really high rank 256. You could make it smaller but it wouldn't retain all the intricate details of the carclueless
>>101939840ponyman went all in on auraflow, "disasterpiece"
>>101939897Before I start, is there anything I should do to run this on a 16gb card. I know people have done it, I just don't know if they did it with this.
>>101939966>All that wasted gpu time.Flux honestly does fine at rank 16, I don't think going further beyond that is going to do much for you than deep fry whatever you want to make.
>>10193998416gb? That's gonna be rough, run it in with the low vram args I guess and probably only train at like 512 resolution (not as bad as you think) and maybe consider lowering the rank to 8.
my puter is making pics of my waifu
>>101939967It says sneed on the poster?
>>101940001I've seen this one already, though it does look nice.
>>101940030>>101940038
>flux is released>every single AI company release a new, much more powerful closed source modelits over
>>101940074>every single AI company release a new, much more powerful closed source modellike what
>>101940030supposed to be topless muscular men, adolf hilter and godzilla. got it right a few times.
>>101940074now the standard is flux, everything inferior will be dismissed, at least that'll force them to work harder and give us actual good products, which is always welcomed
>>101940013Still struggling. Keeps running into unicode string errors. I should just boot into ubuntu and do it there. Don't know if I can be fucked though. Might just wait for kohya
>>101940088Grok2 and Imagen 3 (ImageFX) just got released.
>>101940167That's nice, the genie is definitely going out of the bottle, I'm tired of all that fearmongering of AI
>traditional japanese art style, ink on paper, a cyborg samurai in a futuristic Tokyo with katana and jingasa, red sun, japanese calligraphy on the upper right corner, wabi-sabi, henna and carmine, sepia, minimal brush strokesIThat was was made by flux-pro, can dev reach that level?
>>101940167What was that image made with?
Do we have a baker?
>>101940215Grok2.
>>101940225>Grok2.grok2 is dev pro though?
>>101940225So Flux Pro
>>101940225Now bad, has a more natural look. Let's see how finetuned flux will be like though.
Here we go...>>101940241>>101940241>>101940241
>>101940167>Grok2I'm not worried, Grok 2's image gen already got leaked.
>>101940167>Grok2Flux pro>Imagen 3 Doesn't look any better than Dev
>>101940248>Grok 2's image gen already got leaked.what?
>>101940213I will try and post in the new thread
>>101940167Looks real, did you use anything else?
>>101940257Because Grok is flux, idiot.
>>101940331it's flux pro, and we don't have that retard
>>101940255where is her right hand?
>>101940255>Doesn't look any better than DevHighly disagree. Case in point >>101939483
>>101940372looks like a SDXL gen, the far away details are bad compared to flux
>>101940360she is disabled
>>101940341demonstrate something Pro can do that Dev can't
>>101940442Bro did you even look at the flux images using the same prompt in this very own thread?>>101940442Pro looks much better. It's not even close.
>>101940378I didn't specify for telephoto
>>101940541Come on man that's not a quality picture, that's some ugly mush, flux doesn't fuck up like that
>>101940562that's a but queue anon... try and gen a wren, a baby quoll and a salamander on moss in flux i'll wait...
>>101940629that I agree, flux doesn't know enough concepts, but it's easier to add more concept to a model than giving it more quality in images
>>101940655both have strengths as with any model, hopefully a good finetune will add some more textures and film tropes you can see what they shied away from in training
>>101938755>>10193878628 screencaps....lazy, the right way is run the bluray through PySceneDetect, detect & remove near-duplicate images, caption with joy caption, do finetune on the final 1000~2000 images with random crop enabled, high epochs and slightly slow learning rate, then extract lora at different ranks and take best onequality in, quality slop out
>>101941296>captioning before random crop
