Previous /sdg/ thread : >>100190707>Beginner UI local installFooocus: https://github.com/lllyasviel/fooocusEasyDiffusion: https://easydiffusion.github.io>Local installAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI (Node-based): https://rentry.org/comfyuiAMD GPU: https://rentry.org/sdg-link#amd-gpuIntel GPU: https://rentry.org/sdg-link#intel-gpu>Use a VAE if your images look washed outhttps://rentry.org/sdvae>Auto1111 forksForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeAnapnoe UX: https://github.com/anapnoe/stable-diffusion-webui-uxVladmandic: https://github.com/vladmandic/automatic>Run cloud hosted instancehttps://rentry.org/sdg-link#run-cloud-hosted-instance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restInpainting: https://huggingface.co/spaces/fffiloni/stable-diffusion-inpaintingpixart: https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma>Models, LoRAs & embeddingshttps://civitai.comhttps://huggingface.cohttps://rentry.org/embeddings>Animationhttps://rentry.org/AnimAnonhttps://rentry.org/AnimAnon-AnimDiffhttps://rentry.org/AnimAnon-Deforum >SDXL info & downloadhttps://rentry.org/sdg-link#sdxl>Index of guides and other toolshttps://codeberg.org/tekakutli/neuralnomiconhttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>View and submit GPU performance datahttps://docs.getgrist.com/3mjouqRSdkBY/sdperformancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html>Share image prompt info4chan removes prompt info from images, share them with the following guide/site...https://rentry.org/hdgcbhttps://catbox.moe>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdgOfficial: discord.gg/stablediffusion
https://strawpoll.com/wAg3ArVAWy8/resultsDon't forget to vote.
>>100195251Just like in real elections, I refuse to vote
Can someone help? I want to use Pony Diffusion, but my SD setup refuses to load it. I use AUTOMATIC1111 (I made sure it was up-to-date). Do I have to download SDXL first?Last time I was active here was when LoRAs were still new so I'm a bit behind and the SDXL links in the OP don't really answer this
>>100195368What error is it throwing and when did you last update Automatic1111?
>>100195368>Do I have to download SDXL first?I don't have SDXL and I use plenty of different Pony mixes. Give error msg
>>100195368>Pony DiffusionIs that for anime gens?
>>100195394yes
>>100195394one could probably also gen a black leather couch, sitting against an office wall, if you're into that sort of thing
7900XTX or 4070TI super for my dedicated SD rig7900 is about $100 more expensive
>>100195394>>100195368You should just have to place pony in the "Stable-Diffusion" folder located in Models then select it.>>100195440It's intended for furries/anime but ok.
>>100195440score_9, score_8_up, score_7_up, score_6_up, score_5_up, source_cartoon, rating_explicit, realistic, BREAK front view, 1girl, 5guys, girl on couch, guys standing behind couch, facing viewer, ((loli, little girl, young, size difference, large male, small girl, short, sitting Indian style)), (Laura Kinney), long black hair, green eyes, choker, stockings, abs, slender, nude, (pubic hair:0.8) small breasts, large cocks, dicks out,Piper Perri meme
>>100195463I'm happy with my 4070ti but wouldn't mind more vram... Depends if you can get the XTX to run all your stuff.
>>100195483i'm pretty confident i can get the 7900 to run, i'm using a 6800xt right now and it pretty much worked out of the boxthe 20gb is tempting
>>100195503Then it's pretty much a no brainer. Get the 7900 and that vram will futureproof you.
>>100195368Pony is just an extremely heavy SDXL fine tune. It's unlikely that your Auto1111 is good at all. I recommend you download and use https://github.com/lllyasviel/stable-diffusion-webui-forge which is a straight up upgrade.
>>100195555>Forge>Update last month>Automatic1111>Update 4 days agoAnon, I...
>>100195599Yeah I forgot all the breakthroughs in SD in the PAST WEEK!!!!!Fuck off
>>100195611Bitch, Forge is abandoned.
>>100195620And Auto1111 is complete and utter shit. Things don't need to be autistically updated for no reason. Auto1111 needs constant patching because it barely works.
>braps @ uu're move, homofag
>/sdg/ - where you can practise your snark because life has beaten you down.
How important is BF16 for LORA training?I just found out that the RTX 8000 (48GB but 2080 TI performance) doesn't have that capability.Would NVLinked 3090s work with LORA training? I know NVLink does nothing for image generation itself, but if it works well for training then that would be a cheaper option with more modern hardware.
>>100195628>because it barely worksskill issue
>>100195655>inb4 memory management meme
>>100195651I train Loras at 8-bit and single 3090 will train most Loras in a couple of hours. As far as I know multi-GPU is a massive pain.
>>100195620t: homosexual
>>100195683>meme from last month
>>100195693t. tranny
>>100195686>and single 3090 will train most Loras in a couple of hoursHow many steps? Are all of your training images different sizes?
>>100195710t: ranny
>>100195575food
>>100195719It depends, I'm also subscribed to the multi-crop meme so I only do three buckets and all my images are cropped for it: center, horizontal, and vertical crops aka 768x768, 1024x768, 768x1024. But basically 50-100 epochs (1 repeat per image). Depends on the size of your dataset of course for actual time.
>>100195599auto still sucks for SDXL, though forge being abandoned like that hurts a fuck ton, I tried to manually add the new scheduler thing but forge code was a lot more convoluted on that so I gave up
>>100195805>auto still sucks for SDXLwrong again
>>100195830I tested, it was better but still slower than forge, but I will test again
>Train SDXL lora on base model>It literally doesn't do anything>Train on finetune>lora turns out fine.I don't get it. I'm not even talking about pony fine tunes here. If I train a Lora on the SDXL base model it wont work with things like realistic vision etc.
Is there a place where I can see a list of characters PonyXL recognizes?
>>100195805Well I get my stuff done with it. Forge was crashing with Highres-Fix once too often so I went back to A1111
>>100195847just the og rentry
>>100195848>Forge was crashing with Highres-FixMaybe it's just me, but when forge does Hi-res fix it takes a lot longer and my GPU fan goes crazy trying to meet the demand.
>>100195862Tried tiled VAE?
>Throw random images into images folder.>prodigy optimizer>cosine with restarts = to number of epochs>hit start trainingezpzAnything more than this is snake oil, including captions.
>>100195892>including captionsalmost had me
>>100195892I'll try this tonight.
>>100195897I put that there so someone could actually convince me to take them seriously. The prodigy and cosine with restarts part was real though. I've been having a hard time determining if captions make my outputs worse or not.
im looking for comfyui xy plot functionality beyond what efficiency nodes provides. what are you using right now?
>>100195910Captions are absolutely required. As if PonyXL could've been trained without captions.
>>100195846Probably wrong settings. Have you read through a tutorial first?
>>100195924Yeah, but I'm talking about a LORA here. If the objective of the lora is to help the model produce something that is similar to the input, is there any point in captioning it? The model already understands the concepts being trained into it usually, it's just teaching it to better fit those conceptions to the input images. I feel like captioning if not done 100% perfectly can actually harm the results whereas not captioning let's the model do the heavy lifting of deciding what everything is in relation to what it knows.
>>100195924I thought captions are only the things that you don't want included in the lora
>>100195910prodigy shouldnt be used with restarts, but you probably arent enabling the restarts so its fineby not using captions you burn the training images into the model, making it less flexible and not paying much attention to prompts, captions make it generalizble, it isn't much of a problem if you are generating the same kind of stuff as the training images thought
>>100195938>Probably wrong settingsWhat setting could possible be wrong that causes it to shit the bed on base SDXL and nowhere else? It seems like a fairly specific problem.
>>100195956>but you probably arent enabling the restarts so its fine.Guess again.
>>100195914good gen
>>100195947Try training an upskirt Lora without captions on base SDXL. The AI is not "smart", you need to associated words with the features of the images you feed it.>>100195953You want to associate the new concept you're trying to train with the model's existing knowledge. As the other anon said, this will ultimately make it generalize better.
How would prompt for this pose in pony (pic related)? I don't know what it is called but "side view, all fours" isn't getting it done (creates overly elongated pose even with controlnet).
>>100195996>Try training an upskirt Lora without captions on base SDXLOkay, that is a pretty good argument. Most LORAs I train aren't that specific.
>>100195996I read a tutorial that said "we caption blue dress here otherwise the lora would always generate a blue dress on the character" or something like that.
>>100195980oh, it isnt overcorrecting the lr at the start so its fine, cool
>>100196007>pony>controlnetControlnet sucks with Pony
>>100196007profile view, laying down, butt up/raised
>>100196030thanks I'll give it a try>>100196029I have used it plenty with pony, most recently >>100195023
>>100196007side view for surelying, on-stomach for sureand then something to get her ass in the air. maybe ass up
>>100195947Without captioning, your lora won't be very flexible and would likely be burnt.
>>100196016That sad truth about LORA training is that almost nobody knows what's going on. What helps one dataset might not in another for an entire myriad of reasons.
>>100196016Yeah basically, it's actually surprising what the AI can "learn" sometimes and what things it ends up paying attention to.
>>100196040this would be good too
Still wondering if anything pre-Ampere isn't ideal for LORA training. I read that BF16 is what's used for it and fine-tuning, but I'm not sure how important it is since other stuff also works.Basically, are Volta GPUs and RTX 8000 not ideal for LORAs despite their huge VRAM sizes?
Do you think SD3 will shit the bed with controlnet like it for SDXL?
>>100196053>what's going onI bet that's buried in some paper and everyone just used hours and hours of trial and error until he found something that works somehow
>>100195994>good genThanks. Strategy is pick random past gen or recent post hereabouts, run it through a couple of rounds of img2img with this prompt :style of Frank Frazetta| style of Andre Kohn ] Impasto, 3d acrylic, chiaroscuro , manga , by Bonestell, by Lucien Rudaux, barren surface of the exoplanet , by Frank R. Paul, by Charles Schneeman, Charles Bittinger , space art, by Alberto Morocco , (space art:1.5) by by Dan Matutina,BREAKlifeless surface of exoplanet by Alberto Morrocco ,, , , , amazing, scifi , (minimalism:1.7) surreal, style of Igor Morski, [ style by Rob Gonsalves | style by Dan McPharlin | style by Igor Morski | style by Andrzej Grenda | style by Agostino Arrivabene ], negs: frame, fish, little man, distant man car ,stalagmite, mountain, hat, moon, planet, umbrella, astronaut, reflection
>>100195961The finetunes might be more familiar with your images. Have you tried adjusting the learning rate and number of repetitions?
>>100196085I've seen quite a few of your posts now brudder. I would give up on it. You don't want to be on unique hardware and I'm almost certain you don't want to be on old hardware. The vram is nice but people are training loras on 6gb so don't think about it too hard.The reality is>>100196053
>>100196007Are you using SD1.5? You need to train a custom lora if you want to create that specific pose consistently.
>>100196128
>>100196085VRAM is meaningless if the GPU itself is slow as shit. VRAM is only important up to the point of where you fill it up, basic Loras can be trained on 12 GB of VRAM and probably 4x faster than a dusty old GPU with a 4060ti 16GB.
How low does your loss go when training a LORA? Mine never seems to dip below 0.110.
>>100196144VRAM can greatly affect speed. What should take 30 minutes can take hours if you don't have enough VRAM.
>>100196128Well I don't necessarily need to be able to nail it consistently, it is just a one-off image. I am getting a bit closer to what I am looking for after improving the prompting with anon's help which I'd post if catbox didn't suddenly die on me. I might ultimately just copy and paste the reference image's midriff part into my gen, get the coloring to match in gimp and inpaint over it to get it perfect.
>>100196203VRAM isn't what that chart is measuring, don't want to burst your bubble. The VRAM only holds the images in memory, the processor still does all the work. Your shitty 6 year old GPU is not good because it has a lot of VRAM.
>>100196089thanks for the info anon
>>100196225Yes newer GPUs are better than older ones per gigabyte of ram just because of memory speeds. Nobody would argue that because if it wasn't true, nobody would buy 4090s over 3090s. I think what that particular anon is trying to articulate is that adequate vram is better than no vram at all.
>>100196198That looks normal. Losses don't predict out well a lora would turn out.
>>100196249>Losses don't predict out well a lora would turn outWell, yeah, but they're a pretty good indicator of if something has gone seriously wrong. But the reason I asked the question was because I can train a lora for hours and its loss will slide down to the .12 -.11 range very quickly and remain there regardless of how many more epochs I do. You'd think it would continue to dip albeit slower than before. And just because the loss has stabilized, it doesn't mean the training is complete. Training for longer does improve the output, yet the loss does not change. It's just a little confusing.
>>100196198The loss is almost meaningless as a metric, it's all in the samples. 0 isn't even good, that means it's literally burned the images in. Since the model is generalizing your images hopefully, it should never perfectly recreate your images.
>>100196230>thanks for the info anonNo prob. Rather than go for control and precise style, it's fun to set up SDXL to trip over styles and hallucinate on some features of the prompt.
>>100196225If you don't have enough VRAM, your lora training can become very slow or won't even work, especially if want to training with high resolution images. At that point, the GPU speed won't matter that much.
>>100196317GOOD THING I QUALIFIED THAT HUH
>>100196326You said "GPU does all the work". If you're fine with generating shitty low quality loras, then a low VRAM GPU is fine.
>>100196007all fours, from side,top-down bottom-up, looking at viewer, elbow rest, head rest,bed, on bed, bed sheet
>>100196363>>100196030>>100196040We wouldn't be having this discussion if Emad wasn't a talentless hack to acknowledge the need to accommodate controlnet with all subsequent SD models in order to remain relevant.
>>100196362That's what you chose to read because you're a retard.
>>100196436Are you regretting your 12gb graphics card purchase?
This thread feels unusually helpful and on topic. What's up?
>>100196363nice one
>>100196468No lmao
>>100196499d*b* is working at McDonalds right now
>>100196529I like this one.
>>100196519Why is your 3090 so cool? Mine idles at like 60. Did you power throttle it or something?
>>100196586he lives in an igloo
>>100196519Lmao, a waste of VRAM if you're only interested in generating basic loras.
>>100196598Yeah anon, that's what I'm doing, training basic Loras. You have a really bad habit of making asinine assumptions, you should work on that. I'm sure that's why you don't have any friends.
If anyone is using Sea Art XL, posted a quality improvement negative LORA: https://civitai.com/models/421889Working on training one for AutismMix XL now, let me know if anyone uses any other XL models and wants a general better quality negative LORA.
>>100196598NTA but you gotta stop moving the goalposts.
>>100196574thanks. prompt is [Suprematism}Constructivism|Bauhaus] [style of Igor Morski| rococo |style by Rob Gonsalves | style by Dan McPharlin | style by Igor Morski | style by Andrzej Grenda | style by Agostino Arrivabene|[style by Fischinger| style by Hoyland] cyanotype, infrared photography, (minimalist:1.3) , by Roger Dean, by Chesley Bonestell, light and shadow at https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma
>>100196622Maybe put some plotted comparisons with subjects people normally gen using the seaArt model That's all I got, thanks for the contribution
>>100196623Nta, but he's disproving his own argument with that image. Why did he get a 4090 if he thinks 12gb is enough?
>>100196659Well, you can do more with a 4090 than just train loras and 12gb is enough to train a lora if you're willing to make concessions for time.
>>100196618Yeah, I'm sure a person going ballistic over online posts have a lot of "friends".
>>100196659Because I know for a fact it doesn't use all the VRAM on my 4090?
>>100196705Because you're being an obtuse antagonistic faggot who makes bad faith assumptions about other people through willful bad reading. Be less of a faggot and maybe I'd be nice to you.
>>100196586>>100196594u prolly have shitty air flow
>>100196507tanks
>>100196651people really underuse that style of prompting
>>100196722I appreciate you. You've been helpful and also done your part to shit on retarded trolls. You make sdg a better place.
>>100196499theres only one poster we know of who had to "take time away from the thread" who otherwise squats here for a majority of the dayi dont think this is circumstantial
>>100196731>anon completely misses the point about ambient temperature
>>100196777I like my computer room at a brisk 69 degrees. But my computer does have a lot of fans and good air flow :^)
>>100196709Your basic lora training didn't fill up your 12gb card, so obviously doing the same thing won't fill up your 24gb card.
>>100196754DANNIIIIIIIIIIIIIIIIIIIIIIIIII
>>100196760>people really underuse that style of prompting.On the other hand, going for that precise pose and style produces amazing results.
>>100196722Lmao, you're the one doing it. Look at how many insults you're using.
xl is weird
>>100196873>xl is weirdThe artists works with his tools limits.
test
>>100196731Interesting. I took the PVC window off the side of my PC and it cooled right down. Must have been suffocating it.
>>100196891Aight
>>100196903Also dont use the default meme silent fan profile.Those are retarded, since they have like 0rpm till 55C.Once temp hit 55C the fans rev to cool the card down, and so on and so on. Do a custom curve in afterburner
>>100196889it worked, anon
>>100196889welcome back king
>>100196970wtf anime real
>>100196947I've had them cranked up for a while
>>100196975>>100196970thanks. just checking the wifi :)
girl, drop it to the floor i love the way your booty go all i wanna do is sit back and watch you move
Quick update on that cosine with restarts learning curve with prodigy
>>100197054very cosiney. i like the way it restarts
>>100197063idk if this was intentional or not but I laughed.
>>100197063i see merchants everywhere nowadays
>>100197156>idk if this was intentional or not but I laughed.yeah, there's some parody in it.
>>100197174>i see merchants everywhere nowadays"corporate logo" was somewhere in the prompt
>>100196754boob
>>100197174That's because you're racist.
>>100197204uhm what about them?
>Batch size over 2 reduces lora training qualityfact or cap?
>>100197369Probably not quality but it will impact the learning rate.
>>100197369paranoid schizophrenia but it's a boon in lora training
>>100197366gib
>>100197419ACAB
>>100197369If you're adjusting other parameters in order to increase batch size, then yes.
>>100197429here you go
>>100197451>other parametersfor example?
>>100197462Thank. I guess I shouldn't have assumed the reaching for booba lora was self explanatory.
>>100197369Something something stochastic vs minibatch gradient descent
>>100197325hard not to be these days, unless youre living in a hole
How much israeli influence is in the AI scene? probably alot, I know that the Open AI CEO is a zionist, a violent one at that.
>>100197526it was, i was just making conversation ^^and yer welcome
>>100197539
>>100197557This picture has an interesting subject matter and unique art style. Is it a lora?
>>100197557
>>100197540being israeli doesn't inherently make you a zionist. there's a lot of tech industry in israel but they might have a lower zionist-per-capita compared to the israel average. that said though, there's a fair bit of AI companies in israel. nvidia just bought one this week
>>100197568no, just base pony with hidden artist tokens.https://lite.framacalc.org/4ttgzvd0rx-a6jf.htmlUse "nev,ufr" for that style.nev is for tekuho style and ufr seems to activate shotas (found this myself by accident)
>enter on topic helpful thread >immediately engage with shit posts and he wonders why anon dislikes him
I really should just get this in a repo before something inevitably fucks up>>100197421very cute brown chibbers>>100197539BtB is always a banger
>>100197618Pony must come with a fucking four inch think user manual.It's the only model I know where you can't just type what you want and it pops out. You actually need to know your shit.
>>100197642if you can learn to speak "tag" and install the tag autocomplete it's probably the single easiest model to use. It's the only one that DOES do exactly what you tell it to do.I will admit the new juggernaut model also listens really well though.
my gens are known for their expressiveness and dynamism
>>100197655Why is she doing that?
>>100197665>new juggernautNew juggs took a pretty big fidelity hit in exchange for hearing prompts better imho.
>>100197733Cute!
>>100197641ty. She was voted the cutest cannibal in her village.>>100197733very nice
>>100197753she ate the president's uncle
>>100197733unemployed homo
So have we all jumped on the PAG train yet? It's one of the few things that have come out recently that is just a straight up increase in quality with few drawbacks.
>>100197792whats a PAG train?
>>100197792It's indeed pretty good, helps a lot with making more accurate anatomy, from the x/y/z i've done
>>100197792The quality increase doesn't appear substantial enough to merit the time needed to learn it's quirks but that's my casual observation
>>100197792Imma be honest. I haven't jumped on and I don't know that I need to. Free U has been working for me and I don't know how either one works or if they do similar/different things.
>>100197792I don't know, I installed it. What's the best scale to use it at?
>>100197434stop resisting
>>100197843cfg 3 pag 3 is my happy spot
>>100197792free lunch!(its never free)
>>100197792I have gotten good results from it but it kind of reminds me of free u in that you can improve your gens but the settings are very contingent on your model, prompt, controlnet, and other factors
>>100197860thread schizo
>>100197843The paper recommends a CFG=4 with PAG=3, but different models might benefit for more or less
>>100197860>The double straps in the underwear.Do these actually exist or is it just something that is mass hallucinated by AI.
>>100197866Free U is "solved" I thought.
>>100197871I've always wondered that myself, due to how common they are. maybe its the underwear meta in anime world
>>100197887That is just for those base models, anime models and photorealistic models can get better results with tuning b1 and b2
>>100197712coz "(breast envy:1.2),breast grab" ^^
>Train lora>many of the generations feature a radiator in the bottom right corner>Across multiple attempts>across many models and parameter sizes>the radiator is still there>Look through the dataset>no radiatorWhere is the radiator coming from?
>>100197851freeze, polce!
>>100197951I've had that happen with hats. No images of character wearing a hat in the dataset. lora keeps genning character wearing a hat.
>>100197968It would be fun to see exactly why the AI is doing that but it's not exactly the kind of thing you can crack open and look into.
>>100197920>better resultsidk about that one. Different maybe but I doubt "better." I had been flipping between two different params for pony I found on h but after trying the "official" one I found today I don't think I'll be fidgeting the numbers anymore. I've had incredible consistency. I guess I should try the params one last time to see if any fuck with artists more or less.
>>100197951I've had stuff like that happen, in the end i fixed it by adding more detailed background tags and adding more varied enviroments, nature, scifi, fantasy, etc
>>100197951Hmm I dunno what your dataset and model is like but you get spooky stuff like that sometimes. Long time ago I trained an embedding on 1.5 for a person. It was a woman and there were NO men in the image set. I used elysium 1.5 model as the base but for some reason randomly like 1/4 images would replace my blonde hair blue eyed beauty with a handsome blonde hair blue eyed man. Why? I dunno. I went over and over the tagging but could find no explanation. Just a quirk of the model that that embedding I guess.
>>100197843>>100197852>>100197868Okay, i did cfg 4 pag 3
>>100197743>>100197753ty anons! love both of your stuff!>>100197985this
>>100197999Made your 90's anime look like the 2016 blu ray remaster.
>>100197955we're here to confiscate your drives
>>100197871A simple search on google would've given you the answer. (Yes, they do exist.)
I think my model needs 5% more Retropixie and 5% less of the realistic stuff I have in there, HLL artist styles aren't quite taking as strongly as they shouldpretty happy with it otherwise, does native high res really nicely
>>100198033I sure saw a lot of mens asses when I google it.
I know this isn't sd per say but what is the best upscaler for cartoon like images? I have some older series that I would like to try scale with this but I don't know where to start
did auto ever fix all the memory problems? it's the only reason I swapped to forge to begin with.
>>100198052Is forge getting updates again?
>>100198044ergscan anime
>>100198059No it will never be updated again afaik.
>>100198052I thought forge had the exact same problems.
>>100198052Personally, I just bit the bullet and moved over the comfy. It only really sucks for inpainting.
>>100197486Example is network dimension. If you reduce it to allow for increase in batch size, then quality might suffer.
>>100198087Ah. I've found the difference between 64 and 128 is negligible anyway. No point jumping up to 128 and eating vram when 64 is pretty much the same.
julien is beans
yw for starting the reel human bean joke after that one anon fubbed up, jules
the beeeeeeaniest
>>100198024You have the right to the remains of a silent attorney! If you cannot afford one, tough noogies!
>>100197951If your training images have too many dissimilar features, SD might "average" those dissimilar features into a new object.
>>100198163Yeah the women in my dataset do generally live in filthy apartments.
Why aren't there consumer grade non gaming gpus with larger vram?I feel like they were a thing a long time ago then everything become gamer shit with rgb lighting.
>>100197714it's like Juggernaut was overtrained on 90s TV moms or something, left is Juggernaut, right is latest Zavy Chroma XL, same seed / prompt
I had fun actually discussing stable diffusion for a while, but I get the feeling we're done now.
>>100198206>what are workstation gpus
>>100198206Nvidia has been making Titans since 2013, but they were always basically advertised both at gamers and non-gamers. The original one still had a $999 MSRP in 2013 dollars also, they were not less expensive
I've been dumping my Tay Tay gens on twatterwww.x.com/tayai42069
>>100198296okay...
>>100198052It's very slightly improved, but overall, it just sucks
>>100198246those are roughly the same quality (worse eyes for LONG ear). only thing right has on left is the little arm hairs. prompt is too simple to take advantage of jugX or right didn't give what you prompted for though.
>>100198296bro I've been dumping asuka gens on twitter since 2022, but you don't brag about that shit
>>100198312I'll just keep using my deadend fork then until it stops working or something comes out that I must use. Thank you.
>>100198134fudge the powice
>>100198314I thought the shitty eyes were particularly noticeable on the Juggernaut one mostly. It seems to be a recurring problem with version 10 specifically.
>>100198346They said as much that with version 10 they used a different captioning system to make the model more flexible. I think that was the tradeoff. If you ever used a llm to caption images you might see why. While they generally do a good job of describing the scene, they also add a lot of noise.
>>100198346oh it's noticeable, but it's something detailer is going to fix anyways.
>holding onto forge for dear life wouldn't be me
Throwing in the white flag, Comfy is absolutely infuriating to try and set up and use. The greatest success I had was importing an auto image and using those settings. Comfy WF dot com, guides, and youtube videos were of little help. I tried lads.
Is there any 2D openpose editor working with the latest 1111?
>>100198384how is dragging an image or loading a json hard?
>>100198384Your only mistake was trying to begin with.
>>100198384its okay anon we still love you and your retarded tiny brain
>>100198395>the absolute state of comfy users None of you shitbirds make your own workflows (except for the two mega autists)
>>100198384I've thrown in the towel a few times, but I keep coming back because at the end of the day comfy can just do what the others cant. It's much better supported and easier to develop for.I appreciate the simplicity of A1111 based UIs but ultimately it puts you at the mercy of a developer who might just walk away without warning at any moment.
>>100198384It's alright. Comfy wasn't meant for everyone.
>>100198410I fail to see how that's a bad thing
>>100198135nice
goon night, may you dream of cute soft squishy things.
>>100198420True and trueFor its target audience it does the job fairly great
>>100198410>make your own workflowsMost workflows are already solved. Making your own when a perfectly functional example usually exists within the folder is pointless.
>>100198384It has a default workflow that couldn't be more simple, like the "purple galaxy bottle" one, what were you trying to do exactly
>>100198415Not trying to incite console wars whatsoever. I will not reply to whatever you say so if someone DOES start console wars, it wasn't me. What can comfy do that any other front-end can't? I was under the impression no front end is more capable than any other, only different.
>emtpy latent node >load checkpoint node >conditioning nodes >sampler node >vae en/decode nodes >save image node you need more?
>>100198435a lot I see are stupidly overengineer too, people are always doing things in more steps than is necessary, and going out of there way to use the stupid standalone integer nodes and all that stuff when they don't really need them
>>100198462kewl
>>100198415>can just do what the others cantlmao
>>100198422>nicethanks
>>100198384>buying into the trannyUI memelow-IQ
>>100198496He's not wrong though Any new developments are almost always slapped into comfy while the rest straggle behind - If anything, comfy can hold that over the head of the other frontends.
Can someone explain the whole tranny UI thing? are there actually trannies involved or is it just because trannies use it?
>>100198520lmao1. comfytroons have been spamming "A1111 IS DEAD" since mid 2023, it's hilarious2. there's not a single USEFUL thing you can do with trannyUI that you can do with every other UI3. there is no "new developments"KYS you fucking trannyme loser
>>100198018that's how I thought too
>>100198438Along the lines of this, https://civitai com/articles/4560/upscaling-images-using-multidiffusi with face adetailer pre and after upscale. Most wf I saw had all kinds of jank that would break the whole wf when I tried to turn off parts.
>>100198547>there's not a single USEFUL thing you can do with trannyUI that you can do with every other UII just illustrated one, anon I use automatic btw
>>100198546comfytroons are literal losers with no life that have turned SD into some endless loser e-drama for some reason, it's the only reason they do anything related with SD and have completely ruined these threads>>100198565where?
>>100198546https://en.wikipedia.org/wiki/Node_graph_architecture
use this autism sampler blend for maximum quality especially at high resolution and/or deep shrinkkarras scheduler. 9 steps euler_smea_dy, 21 steps dpm++ 3m sde
>>100198577>completely ruined these threadsIt's not their choice of UI, anon, it's the fact they use these threads as their personal shitting grounds.
>>100198546It's inane drama between two guys. You can disregard it.
>>100198591both things are relatednot all retards use comfy, but all comfytroons are worthless retards
>>100198597you can't, because for more than a year every single time anyone ask anything the literal answer every time is LOL A1111 IS SHIT BRO, YOU NEED TO INSTALL TRANNYUI HERE LET ME GIVE YOU THE LINKS BRO LMAOand that objectively ruined these threads
>>100198296Oh no. That's illegal in Missouri. A bunch of anons on /aco/ shit themselves and deleted their twitters and stopped posting because the taylor swift ai scandal. Be careful anon.
>>100198642She ain't gonna do shit.You're more likely to get pinched for openly being anti-israel.
>>100198438But if you want to use loras and upscalers and stuff you have to figure out how to hook them up to be desu
>>100198652utimateSD upscaler for comfy UI is completely braindead and plug and play.
>>100198652let them play with their digital plumber simulator
I shitted and I farted
>>100198702Well go on, report on the smell
>Person in lora has a bung eyes>forgot I had mirror images argument turned on>Now both eyes are bung eyesWhy didn't I get two good eyes?
>>100198755With or without PAG?
>>100198759>PAG?
>>100198734issue skill>>100198759is pagged
>>100198766PAG stands for Pictures Are Gooder. Consider using it.
>>100198755nice style, anime and early Hanna Barbera
>>100198773
>>100198799Is that a lora, checkpoint or interface?
>>100198799Incorrect, it stands for Prompters are Gooning
>>100198830https://github.com/pamparamm/sd-perturbed-attention
>>100196970moar? this is actually tripping my brain. im sober
Where the fuck is the SDXL inpaint controlnet?
>>100198950Cool!
>>100198884I liked the anime real version better
Cute ayy lmao
>>100198962
>>100198800Thanks. There is some Urusei Yatsura style lora in there too. They are actually Lum characters. >>100198806vanilla pudding pop yummy
>>100198886they proof they tried to hide
>>100199115Yes, a space knight monkey
>>100198886is that real?
>we're going to the /sdg/ thread again, aren't we
>>100199162all the pictures posted here are real
>>100199162Of course, when has anon ever posted fakes?
Nom
>>100198830Perturbed Attention Guidancegod damn shitters, do better sdg
>>100199183Posting real ayys in an ai thread would be good plausible deniability. This picture reminds me of a nightmare I once had. and that old tv show Monsters. >>100199196I don't know, so I guess never. Nice gen.
>>100199231Is a link to the github not sufficient?
>>100199246I got you
>>100199207Yum
>>100199260I don't know what frontend you're using bud
>>100199231What is it perturbed about though?
>>100199260depends on the level of laziness
>>100199277To be honest I was pretty perturbed no one answered that anon correctly
>>100199267fake and gay. but maybe real.
>>100199275No, you were crying people were making up acronyms for PAG but there was ALSO a link to the github for PAG. People also Good.
>>100199290lol
>>100199294>People also GoodFairI suck cocks
>>100199246Thanks! I love your gens
my most favoritist part of sdg is seeing new ideas by new anons brought forth and being on page 7 with twenty images to go
Penis and Gagina
>>100199353Talking about that, would using some NSFW focused model as a refiner improve the quality?
"Daddy"
>>100199344Aliens if they goon.
>>100198806that would be even cooler without the "DOS game" look
>>100199371Sure why not, knock yourself up
>>100199304is this the disclosure /x/ promised me?>>100199331Thank. Very cool lum!
>>100199304the truth is out there
>>100199441undeniable proof, unexplained phenomena, (UFO:0.1), far away, (the rule of thirds:1.2), UHD <lora:RMSDXL_Darkness_Cinema:0.5> <lora:boringRealism_primaryV3:0.8>Negative prompt: grid, lattice, grainy, noisy, mundane, dull, vague, text, 404, error, logo, drawing, painting, chart, label, cover, soulless gaze, uncanny, staring, close up, upclose, zoomed inSteps: 7, Sampler: DPM++ SDE Karras, CFG scale: 2.5, Seed: 344448045, Size: 1024x576, Model hash: 4496b36d48, Model: dreamshaperXL_v21TurboDPMSDE, VAE hash: 235745af8d, VAE: sdxl_vae.safetensors, Token merging ratio: 0.5, Hypertile VAE: True, Hypertile VAE swap size: 5, FreeU Stages: "[{\"backbone_factor\": 1.1, \"skip_factor\": 0.6}, {\"backbone_factor\": 1.2, \"skip_factor\": 0.4}]", FreeU Schedule: "0.0, 1.0, 0.0", FreeU Version: 2, CADS Active: True, CADS Tau 1: 0.6, CADS Tau 2: 0.9, CADS Noise Scale: 0.27, CADS Mixing Factor: 0.9, CADS Rescale: True, CADS Apply To Hires. Fix: False, Lora hashes: "RMSDXL_Darkness_Cinema: 246490f20190, boringRealism_primaryV3: 23c9fa12f1fd", Discard penultimate sigma: True, Version: v1.8.0
>>100199471None of those people even notice because they don't think to look up. If you ever need to hide, hide in the sky.
>>100199471what does 404 in the negative do?
>>100199518i dont know, i assume filters out 404 error pages and related images
>>100199431is this better? ( ω *)
>>100199535Were you using a lora for that style?
>>100199550All from https://civitai.com/models/420049, it's finetuned from autismix, so a lot of styles also work
Theres debo right on schedule
>>100199676
>>100199561Looks really good. Some people might not download it because of the file size though if there's any way to prune it. But I don't know if that's possible, loras are the most I can do with my gpu.
>>100195480you forgot (dark skinned male:1.5)among a few other necessary tags
B R E A DUNNY
>>100195480LK anon is back? Haven't seen you in a while. >Piper Perri memeThere's a 1.5 lora for it, but I haven't seen a pony one
>>100199706Thank you, just noticed it defaulted to show the fp32 full model in the download button, should be fixed now ≽^•⩊•^≼
New>>100199595>>100199595>>100199595Thread
With comfy you build the workflow With auto you are the workflow
