Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101601667>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/ComfyUIInvokeAI: 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://www.modelscope.cn/homehttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>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>Kolorshttps://gokaygokay-kolors.hf.spaceNodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper>AuraFlowhttps://fal.ai/models/fal-ai/aura-flowhttps://huggingface.co/fal/AuraFlows>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>View and submit GPU performance datahttps://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>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
blessed thread of frenship
bigma will keep us save from the space pee aliens
official pixart bigma and lumina 2 and also the hunyuan finetune waiting room
Bigma status?
time to start genning some shit.
>>101616817stuck in 256 x 256 latent space
>>101616767Thank you baker for providing us with fresh bread >>101616777And thank you blesser of the bread for saving us from the evils of the latent world
>>101616867this is where the monsters live anon....
Jungle time I think, sayt anything you like i will gen but has to be in the jungle, imagine your place is arriving towards the crude land sight.
>>101616925try asking the latent space for a "1girl" i want to see what it does
>>101616941ok, i was having a real problem just there with it not genning 1girl... but you make it easy for me now anon i can give you 1girl in jungle no problem for now. But I'd like to control it.
>>101616941She is not yet detailed, i'm still setting up.
>>101616941>"1girl" i want to see what it doesdepends on what you want? younger i'm doing that here even if fully clothed so fuck off.
>>101616864lets see it
>>101616993pp samplers are interesting
>>101617012They are indeed. Also deis and the new beta schedulerGood night anon!
>>101617012no i hate them for most cases, ddpm works
>>101617078WAAAAAAAAHHHHHHHHHHHHH!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
>>101617078maybe lower cfg will work, be right back.
0.4 cfgeuler_cfg_ppinteresting but not entire to spec.
>>101617108hmm time to try with pony model at 1024
>>101617049>beta schedulerlooks incredibly similar to simple and sgm_uniform mixed 50/50 with exponential 50/50. im not sure what alpha and beta values do
>>101617130was already pony, just forgot to up res, the resulting image is a naked women though and i'd get banned, its decent mind you.
>>101617144putting "nsfw, nipples" in the negatives will usually net you child friendly images
>>101617155i don't care, why care? I',m a fucking man i don't care what these idiots say...
>>101617151WWWWAAAAAAAAAAAAAAAAAAAAHHHHHHHHHHHHHHHHHHHHHH!!!!!!!!!!!!!!!!!!!!!!!!!
bye anons, good night.
>>101617174gn
>>101616817in latent space
>>101616867>>101617296Hype tho
>>101616767can someone make me a picture of a
>>101617308>an astronaut is riding a brown horse on the moon, in the background is the planet Earth, double exposureone day it'll be the whole prompt
>pos: zebra pattern, ...>neg: animal face horse, ...
>>101617570this is cool
>>101617575nice patchwork
>>101618090i don't like the way it's looking at me
>>101618090I really like the way it's looking at me. Catbox?
i sleep till bimimimimigma... zzzzzzzzzzz.... mimimimimi.....
mabig
>>101617575Little big planet
>>101618928Hansel and spagrettel
cough
>>101616767cool gens
Good morning anon>>101619135Nice house
>>101620339gm
>>101620383gm
Closer
>>101620502this one is cool
>>101620502ty
MFW I was born with too many fingers
>>101616767I got me a question for everyone of the thread. I do both local text chatting and image gen, and will be building a full blown desktop for it. I have the money for the 4090, but part of me is curious about the jump between the 4080 super and the 4090.Keep cost out of it, that's not the important part. Would the 4080 super be the better choice? Part of me feels that way.Also, is there a big difference between the ryzen and Intel on generation? I know trying to image gen on an amd GPU is a pain in the cock and I won't be doing that.
>>101620572>4080 super be the better choicenooope, vram is king here, especially in text gen. and i think unless you're planning on training/finetuning your own models a 3090 instead of a 4090 would be a better deal.
>>10162057216GB vs 24GB.. you always need more VRAM but 4090 is still expensive.
>>101620603Unfortunately I live in SEA and the cost of a 3090 is just as high as a 4090. Does the 4090 still have the melting port issues?>>101620610Not even the price being an issue (not trying to brag, I saved up all my funny money specifically for a badass beast of a PC). I originally wanted to *just wait* for the next iteration to drop, but unfortunately being in SEA the first run of 5090s gunna be double the price due to scalping. I am not paying nvida yacht tax and some jackasses scalp tax because he fucks the ladyboy doing inventory
>>101620657Also, to contribute, here's something I made.
>>101620657>4090 still have the melting port issuesPower-limit and you won't have to worry. The connector is only designed for so many inserts and running full tilt is what gets you in the fire zone. I set mine to 330w>sudo nvidia-smi -pl 330>saved up all my funny money specifically for a badass beast of a PCThat's cool. You have time to wait for sales then! Buy it piecemeal and set price alerts
>>101620657>Does the 4090 still have the melting port issues?sorry i have no clue. i think a 16gb card should be fine for image gen but from what i've heard 16gb is kind of useless in text gen due to the available model sizes. basically any model you can fully run off your 16gb card you can do so with a 12gb card. maybe it might be handy if you want to run an llm and a image gen model at the same time.
>>101620684Friend, I've done everything including dropping bribes to shop owners to alert me first at sales of the PC parts. The Asian market (moreso sea) is absolutely brutal towards anything electronic.That being said, knowing about the 4090 and how to handle it, I'll probably grab that.>>101620692Well, with a kobold+GPU setup, I can run 20b models at rather nice speeds. And with 64gh of ram, plus the 16 or 24 from the GPU I choose, it would be more than enough to suit my needs. Especially with the new IMAT models, which are 25 percent smaller. That being said I'm an amateur at all this and this may be entirely wrong
>>101620743Seems like the best move. Good luck anon
>>101620743>https://huggingface.co/spaces/NyxKrage/LLM-Model-VRAM-Calculatoryou can use that to see if the quant of the model you're using will fully fit into your gpu. keep in mind that when you offload the model to cpu speed goes down drastically till hits pure ram speeds.>That being said, knowing about the 4090 and how to handle it, I'll probably grab that.i think this is the best move as well
>>101620787>>101620797Thanks frens.
>>101620826this is cool, prompt?
>>101620834"A bronze sculpture of a decaying hand reaching upward, fingers outstretched towards the heavens, as moss and small flowers sprout from its open palm. It symbolizes the impermanence of life and the enduring power of nature."
Work came early today. I'll be back and forthBreathe in, gen out
Screw it, one more before I go
>>101621357>>101621382have fun at work anon, i'll cough here in the mean time to keep the thread nice and warm
can someone interpret for me what the fuck the emotes mean on civitai?why are people crying at Misty? are they uohhh'ing ?is the crying face actually the downvote button?if so, what's the laughing face? it can't be the upvote button because there's already a thumbs up emote.this is fucking stupid
>>101621466i have zero clue honestly. clicking an emote gives you buzz, so i assume it's just people clicking whatever they think looks the funniest.
Sneak gen before meeting starts
>>101621400ty anon, have a good day yourself
>>101620517Water should be more blue low left corner. Great pic
>>101620445v1.0? 0.o
>>101621207Cool.
>>101622464
seems like the model checkpoints that have "mix" in the title are usually better than others. probably because the person making it has already identified the best models that they're mixing. they've obviously done a lot of comparisons.
>>101623005kek good one anon
>>101623032i know it sounds like they're just amateurs mixing shit together, that's what i thought, but if you think about the kind of person who's mixing things together they'd be the kind of person who is doing comparisons.
>>101623005merges > mixes
>>101622002ty>>1016224552k validation after another epoch>>101622962Nice
>>101623355>coughWhen sickness heals the thread
>>101623628that's the phlegm i coughed out
>>101623661The feeling of clarity
>>101618133stares at u artistically>>101618199i dunno what that means
first time it actually did an astronaut and horse shaped blob
>>101623852Yeah, I can kind of see it.
>>101623852>>101623950LGTM
So did any of these new models crack the hot tub full of sausages problem yet or are we still at babby tier levels of comprehension?
>>101624166Image models will not get perfect until they progress from color blob hallucinations. AI will need to be able to construct the scene in 3D space before it can have great comprehension.
>>101623793>i dunno what that meanshe wants you to upload your image to catbox.moe and post the link here so he can peek the metadata
>>101624166Prompt?
>>101624658Really wants text when it sees quotes. Sorry didn't notice
I can never decide between 20, 25, or 30 steps, or even 35. Has anyone scientifically determined the diminishing returns?(typically use euler_a or euler_smea_dy)
>>101624757Good question. I'd like to see a plot of fidelity vs steps
>>101624686nice
Is there a comfyui method to save the result multiple times as it's generating?Instead of needing to regenerate the whole image again for different step counts?
>>101625122you can chain samplers to do exactly that
>>101625122From what I understand the steps is the model's goal point, step 10 for 20 steps is not the same step 10 for 40 steps. It's like telling someone to draw something in 20 minutes vs telling someone to draw something in 40 minutes.
>>101625171>>101625220if you chain 20 samplers together that each do 1 step, is that the same as 1 sampler that does 20 steps?i'll be testing it myself shortly
>>101625252The only accurate way to test is running each step count separately. I'm assuming chaining is just running steps on top of a finished image. Like I said, the denoising schedule is dependent on the target end step.
>>101625252If you correctly set each nth sampler to do just the nth step of 20 it should be the same>>101625289> I'm assuming chaining is just running steps on top of a finished imageIt is not, the advanced ksample node lets you control all that.
>>101625298Then you're just short cutting running multiple denoising schedules in parallel and the time to do your tests would be the same as doing three separate images.
>>101625317what? no.It is sequential and lets you grab the latents at step 20/30/etcthree separate images means doing steps 0 to 10 three times
>>101625332I ALREADY TOLD YOU THE NOISE SCHEDULE IS DEPENDANT ON THE TARGET STEP COUNT
>>101625356YOU CAN DEFINE THE STEPS TO RUN SEPARATELY FROM THE TOTAL STEPS WITH THE ADVANCED KSAMPLER NODEDON'T DOUBT ME EVER AGAIN
>>101625414I'm telling you you're a fucking moron.Which step is 50% of 20Which step is 50% of 40You do know how denoising works, right?
>>101625433You think the anon cares about that? He just wants the image at step 20/30/etc.
>>101624489Trippy texture
>>101625811>>101625875Very cool
>>101625875
>>101625811>>101625875nice
>>101624166did >>101624250 pass?
>>101625968no, it's full of water
>>101625883ty>>101625911Ghost fren>>101625914ty>>101625968>>101626017Seems like this is one of those English being vague things. "entirely filled" != "full of" as something "full of sugar" is not 100% sugar.It does struggle specifically with stuff like making an Orange a different color (if you can guess why).
>>101626017>>101626314kek "entirely filled" just made the hot tub a sausage. Fail
>>101626314Let's see the tub filled 100% with 100% real sausages.
>>101626418how about an empty hot tub
>>101626427>hot tub filled 100% with 100% real sausages
>>101626444>an empty hot tub"Empty" can also mean without patrons, trying out "dry hot tub"
>>101626418Water in the negatives?
>>101626496Didn't want to cheat, but I'll try it. A dry hot tub everyone
>>101626496Did you want blue foam? That's how you get blue foam
>>101626715
>>101626796
>>101626796>>101626830Last
Later
>>101626913Good gens
>open /ldg/>post single gen>leaveWhy?
>>101627167Why do anything
serious question, what's the best method for setup and os for automatic1111? took me a while to get the shit running on ubuntu 2204, cuda 12.1 , then tried to setup on debian 12, cuda 11.8 (no xformers available) and the it/s dropped like 15% , don't want to download 15gigs in python libraries everytime for a new install, is nvidia base docker image the way to go ? I am at a point where I am thinking of just creating a separate install only for this and then use clonezilla to create a disk image to never have to setup this shit again....
>>101627406if you want current kernels/graphics stacks you might be better off with a rolling release distro.
why wouldn't you just do a tub and then inpaint so it looks more like a hot tub? Is this more than a flex or is there some purpose here? >>101627406new upgrades are needed and you can't dodge the pain of the update cycle. cuda 11 is slower than cuda 12. If you try to use a docker image then you are just going to have to manage more.
>>101627406arch and learn to use venv>>101627658>inpainti think that anon would consider it "cheating"
>>101627688so anon is just finding weakness in a product, bitching it doesn't cover the edge case and claiming massive changes are needed. I was hoping for something fun. I'll continue with my life.
>>101627688>learn to use venvhow risky is it to use that and fuck up your system due to a mistake?Is it safer to put all the ai stuff in a container?
>>101627853venv is a folder, it runs from the folder, it's basically a container
>>101627869I mean, how easy is to to call something outside the venv that should have been in the venv.But maybe I'm overthinking it.
>>101623005>>101623049As >>101623115 alludes to, block merges imply greater quality considering one has much more control than simple mixing. Both creators likely spend equal time comparing however. Block merging is the way to go for a multitude of reasons.
>>101621466It's contextual but sometimes random click too for buzz points
>>101627936well if you're afraid that a python file is going to read your illegal porn then no, venvs are not safebut if your afraid of your dependencies getting mixed up, yes venvs are safe
>>101627999I'm afraid to accidentally pip install things into my main system. I'd like to keep it clean in that regard.
>>101628035When you activate the venv from bin it stays there. I've never had a problem with it mixing dependencies.
>>101625122My thought is by adding ksampler advanced nodes manually and set start/end according to your need.Node1 : 40 steps, start 1, end 21Node2 : 40 steps, start 21, end 31Node3 : 40 steps, start 31, end 40Must use exactly same sampler and scheduler.
>>101628060Will have to read it up I guess.
>>101628082The basics are simple, when you make a venv it makes a folder with a portable python version you used to make the venv. To use a venv you activate it (/venv/bin/activate). When you do any pip stuff it puts it in /venv/blah//lib or whatever.
>>101628114thx for your input
Babe wake up, CogVLM2 just got releasedhttps://github.com/THUDM/CogVLM2
>>101628061depending on the sampler and scheduler, this does nothing.What do you think this is doing?
>>101628551Did you chain it? Connect latent from node1 to node2, node2 to node3
>>101628227Meh, still worse than GPT4Vhttp://cogvlm2-online.cogviewai.cn:7861/
>>101627001ty>>101627167It's called "The one and done">>101627406The best way to go is arch linux with podman (same as docker without bs) and the nvidia container runtine. You can run your docker container and map your models over with -v /localpath:/dockerpath>>101627658We were testing if the model was coherent in that fashion without trickery. One shot straight out of the model is still not feasible for a hot tub filled with sausages instead of water.
>>101628227thx for news>>101628666thx for testing
>>101628621yup. it is the same thing. If it isn't then your sampler/scheduler choice is doing something or settings are just wrong. If you are noise injecting at every 10 steps then I will ask again what you are attempting to do. https://litter.catbox.moe/t5zv5n.png
>>101628666The GPT4V caption seems kinda short to me, but it is more correct. Good details about the clothing, etc. in CogVLM2. Sigma's a 300 token model
>>101626913
>>101626913>>101629046
>>101629028Your workflow only shows image from last node which is the 40 steps. Just add preview/save image node on node1 and node2.
>>101623170How many epochs total?
>>101629143This is 5 epochs. I think training with two sets of captions gives it way better knowledge when doing multiple epochs. The official trainer picks between two by default and I have both filled to the brim (300 tokens) from two separate VLM's. I'll continue training until it starts validating worse, which could be a while.
>>101629115I don't want intermediate trash. The entire workflow is there. If you don't like the result there is nothing I can do for you at this point.
>>101629267I think you're replying to wrong person. The guy asked how save image per specific steps. So I gave him idea to use multiple ksampler modes.
>>101629317I do have the thread confusion. Sorry
>>101629246>>101628860
Apparently the "lore" if you will is t hat I want to kill myself at work or whatever? Where the fuck did that come from?
>>101621777Prompt?
I'm surprised there's not a custom ksampler that allows you to define an arbitrary number of step counts and then once a given is reached, outputs the image. I don't wish to chain 50 samplers in order to see the denoising progression.
>>101630367A charcoal drawing on a charred canvas depicts a solitary figure walking away from a burning city, their silhouette fading into the smoke and ash. It symbolizes resilience in the face of destruction and the cyclical nature of life and death.
>>101621777>>101630367>>101630504Wide
>>101630551Narrow
>>101630572Simple and clean, but complex and messy at the same time. Very nice
>>101630335who are you?
Sleep
>>101620838Nice
>>101631095"Centauranon", apparently
https://civitai.com/models/566526/kolorsI feel like Kolors is the only model that is really great at photorealism, remind me of Midjourney a bit, wonder why people sleep on it
>>101632142>wonder why people sleep on itLikely the "its architecture is outdated" meme. I presume anons understanding is that it's merely a hypertuned XL.
>>101632230desu if they made the same training on a DiT model it would've gotten an insane result yeah, and also the fact that kolors fucking sucks at prompt comprehension because it has been trained with chink doesn't help either
>>101632249>chinaspeakI don't necessarily mind that, indeed the thought of what a buger could pull from semi-chinese latent space is intriguing. I'll give it a shot if and when I have the urge to use XL.
awesome another Laura gen gets into the collage
>>101632142https://civitai.com/images/21848019kolors wins this one, picrel is pixart
>>101620838>>101631475memento mori
>>101632142>wonder why people sleep on itfor me because it doesn't work on 6GB vram
>>101630456Isn't this a feature in auto, generation preview? If you're expecting legit use cases in comfy instead of pure autism you're expecting too much. Comfy provides the illusion of choice.
Newbie here, Why I'm getting this results? Model Juggernaut XL with the requirements at civitai.
>>101634730>Why I'm getting this results?*Why am I getting these results?Low resolution, also a strange resolution. You want a resolution that adds up to 2048, the most common being a square of 1024 by 1024.That resolution will give you better results, though not incredible.For most good diffusions you would take that image and upscale it before sampling it again with very low noise to increase detail.
>>101631475ty>>101631659>>101631751>>101631895>>101632299MFW I ran out of air>>101634730Are you using a fancy VAE or under 10 steps?
>>101635122 plus >>101635141
>>101635141this is sick, prompt?
>>101635168Mixed media made of straws and colorful paper of a house on a sunny day
>>101634984Sorry for the misspelling. Thanks for the tip. I will try with another Sampler with less noise and change the resolution.
>>101635255ty, you always impress me with the sheer variety of your gens. keep it up anon.
>>101635374ty. It would just be spam if they weren't interesting!
>>1016350218 steps, is the recommended for hyper version. Don't know how change VAEs.
been a hot minute since I played with SD
>>101630335>>101631777Hello Centauranon. Let's skip the lore of you killing yourself. You're going to die anyways. Why not let it be a surprise how? Plus, who is going to make the centaurs if you check out early?>>101635533What do you use?
>>101635533
>>10163559650/50 merge of AbyssOrangeMix2 and LoliDiffusion20 steps on euler a and a 2 step ultimate sd upscale>that smirk
>>101635517>8 steps, is the recommended for hyper version.Yep, that's fine. Other models you need more steps. I've seen that kind of degradation from custom VAE, poor scheduler+sampler combo (GPU samplers usually), too low of steps on a normal model, and from >>101634984 >Low resolution, also a strange resolutionComfyUI is complicated at first but keeps you in the driver's seat. https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#installing>>101635674> Base Model SD 1.5So you mean you're back, and not that you're using something else. Mistook it for a tuned Kolors gen. You should try that one out if you haven't
>>101635729 (You)Forgot pic
has anyone heard of noise playing while generating?got a new pair of headphones (sennheisers) since my old ones broke and now I notice when I generate stuff there's this weird quiet white noise that plays through the headset. Whatever's going on it's not through the PC since Audacity doesn't pick it up. It's really weird.
>>101635729Just been busy the last couple months troubleshooting my pc.Turns out my pc is fine and this issue I am having is happening on every pc I have tested, all with Nvidia gpus.So just taking a breather.Also got sick of having 6 or more computers in pieces cluttering up my room.>tuned Kolors genI will check it out, thanks.I am assuming I need to download all of it and not just the safetensors file? (On HF btw)
>>101635805Let the latents speak. Also that probably shouldn't be happening. Are you corded?>>101635956Odd.. 6 computers.. have you tried arch linux and docker containers (with podman)?>I am assuming I need to download all of it and not just the safetensors file? (On HF btw)Yeah the ChatGLM3 model is necessary as well as the SDXL VAEFor comfy support until it's native https://github.com/kijai/ComfyUI-KwaiKolorsWrapperAnd hunyuandit is native in comfyui now but you 100% need to translate to Chinese at it
>>101636031>Are you corded?yeah. Difference is my previous headset went through USB while this one goes straight into the audio port so I assume something's going on there
>>101635956And for even more spice in the pot.. all my recent gens are from Pixart Sigma
>>101636031>have you tried arch linux and docker containers (with podman)?no, but I have done a test in Mint 21.1 with gpu passthrough and a windows vm and the issue remains, so I know that it has something to do with the Nvidia drivers.It's definitely a software problem, got a couple experiments to try but that's for tomorrow me to deal with.
>>101634960bless you
>>101628227>just got released what fucking rock you been living under?CogVLM1 was good because it wasn't trained on GPT4 slop. 2 is trained on endless amounts of slop, like every other model.
>>101635729Thank againI'm working with ComfyUI and there are some improves.Swap the model for RealVisXL and add some changes at the workflow. As you say.
>>101634984Nice.
>>101637628yw anon. Nice job. Here's some inpainting examples for more advanced stuff https://comfyanonymous.github.io/ComfyUI_examples/inpaint/
>>101637889oh this is amazing, Now I know how some IA influencers stay at real world. Thanks
>>101638103yw, keep it up and posts some gens!
how many images do i need to make a pony lora? i am giving up on trying to recreate my 1.5 style, i need to use my 1.5 generations to make a lora for pony
bigma
Even though it's 100 degrees outside it's never too hot to bake some new...>>101639278>>101639278>>101639278
>>101639309ty baker!
Filling thread
Full
final cough