Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101857264>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>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>mfw
>/sdg/ is completely deadAnd that's a good thing
>>101861834>/sdg/ is completely deadbased, long live /ldg/
>>101861882long live /ldg/
>>101861881Found the artist, Tachibana Omina
>>101861991w-what is she referring to, anon?
can flux do nsfw
ah shit posted my gen in the old thread
>>101862100No
>>101862108>left : earth in 2024>right : earth in 2030t. climate "scientist"
>tfw classmates, teachers and people around the world have tons of my deepfakes
As someone who has a waifu Dalle 3 kept making into 2B on half my gens there, I'm actually thankful in this case that this model is not overcooked on 2B. Not every girl with white bob cut hair and hair covering one eye is 2B, but Dalle thinks differently half the time. Only issue though is that Flux doesn't have a good vocabulary for exact hair sub-types so I can't get the hair style to be perfectly accurate, and thus I stopped genning. Maybe I'll cook a lora one day.
>As someone who has a waifu>Not every girl with white bob cut hair and hair covering one eye is 2BAre we talking about an OC, or an existing character?
>>101862193do you also rage at 3DPD that cosplay as 2B cuz they arent from the game and their costume isnt 1:1?
Installing flux nf4 on my work laptop with a 4060. What are the chances it works?
>>101862511Damn nice.
Holy shit, maybe we can recreate the Kino from DynamicThreshold + CFG 6 + GuidanceNeg 10 with this one:ToneMap + CFG 10 + multiplier 0.3 (still not optimal though)https://imgsli.com/Mjg3MDgxhttps://files.catbox.moe/1d2u6o.png
>>101862193Well, I think of her as an OC, but her visual design is mostly inspired by and close to an existing character (looks like Hamakaze from Kancolle but not exactly in certain aspects). Flux also doesn't know that character though. It also doesn't know Mashu which would be the next closest character, though her hair cut is yet more subtly a bit off from what I want.
>>101861834>Guys you're ruining this thread with your overt avatar fagging and off topic personal life discussions. It's extremely off putting to everyone coming into these threads who isn't in your weird clique.>"Teehee this is an image generation thread and I am generating images and there's nothing you can do."I'm glad to see the thread dying. They won't even notice the threads have died though. They'll just post quokkas and purple witches and talk about their gender reassignment surgery oblivious that they're completely alone.
>>101862688wood
Not avatarfagging btw. Just seeing how much IP can be found with very specific prompting. This is the last Bulma.
>>101862754very cool picture
goodnight /ldg/
https://new.reddit.com/r/StableDiffusion/comments/1eqs7sq/some_fun_90s_anime_style_gens_flux_anime_lora/Don't sleep on this lora, it can create amazing anime style pictures:Workflow: https://files.catbox.moe/svkcy0.pngCivitai: https://civitai.com/models/640247/mjanimefluxlora?modelVersionId=716064
>>101862769I say that and ofc I get a pretty good one. Prompt is:> Identify the illustration depicting the woman whose appearance and likeness most clearly and convincingly mirrors the characteristics and features of Bulma from the anime series Dragon Ball Z.Wonder if this works for other anime.
What are the options for non-censored? Still only SD1.5?
>>101863049Pony/pdxl would be the other major one.
>>101863049SDXL is way beyond SD 1.5 at this point, especially PonyOnly reason to use SD 1.5 is if you have a potato PC and for some reason can't run SDXL
Is there someone who announced that he'll make a finetune out of flux yet?
>>101863127Multiple. They're all shit smalltime stuff though.
>>101863138can you give me some links? like they posted that announcment on twitter or something?
>>101862702How do you get this style?
>>101863143Literal whos on reddit. I don't care enough to even look them up.
>>101863361>tumor nipples
>>101863372Actually it ass is in place of pussy .
why thread suddenly so slow?
>>101863507I opened it :(
>>101863507hype cooled down, it happens with every new ai tool
>>101863507bedtime for americans
Oh yeah baby fluxing shit up on my laptop
>>101863616First non-test gen. Needs work
https://reddit.com/r/StableDiffusion/comments/1eq98ca/euler_cfg_actually_works_on_flux/What? I didn't know you could use the cfg++ samplers on flux? Everytime I tried euler_cfg_pp got some completely fucked up outputs
>>101862368should work fine desu
>>101863707It does indeed work fine. Genning on the train ride home
>>101863080>>10186307813700k/RTX4080. Last time I used SDXL it was alright. I did 4k+ with SD1.5, and using tricks to start with a higher res gen helped hands and some features.. still limited but I loved DareLites Fantasy Mix.SDXL I found more limiting.. I guess too much stuff removed, it knew of some things better but worse in other areas. Has SDXL been 'fixed' by third party models now?Is Flux1 uncensored, or at least doesn't have as much training data removed?SD3 also cucked?
>>101863765From my light lurkingSD3 is the most cucked model to dateFlux can do ass and titty and vagoo, but I hear peen needs some work.
>>101863805>Flux can do ass and titty and vagooIt really can't though, besides erotic stuff where everything is covered
>>101863820I've literally seen it do perfect pussy lips idk what to tell you anon
>>101863805>titty and vagoo, but I hear peen needs some work.It's the opposite.
>>101863831>>101863829
When I try to load this lora I got a OOM error, https://civitai.com/models/640247/mjanimefluxlora?modelVersionId=716064what the fuck? I have a 24gb vram card and without a lora it's usually at 13gb, why does it ask for so much?
>>101863507it's snore o'clock, back to shleep until the next big thing.
>>101863639I'm not a big fan of euler cfg++, this shit makes things blurry for no reasonhttps://imgsli.com/Mjg3MDk5
>>101863852bro same, keep getting OOM with using one GPU when adding the Lora.I have 48GB in total and I'm currently trying to use the second GPU too but Comfy is a bitch when it comes to that
>>101864032For sure
>>101864264It's weird for me too. It will oom on one generation, I'll stop and try again and it won't oom on the next. I don't know why. I actually think it might be a bug.
>>101864329After yesterday's update I get a guaranteed OOM on the first run, but all the following ones work just fine.
>>101864343>>101864329maybe because of that commit?https://github.com/comfyanonymous/ComfyUI/commit/517f4a94e4a5c45edc64594d70585ec8aeb787e0
https://xcancel.com/Lykon4072/status/1823094103862558893#m>Oh wow look at me, SD3.1 can do women lying on grass now, please hype for us!!
>>101864591>add "lying on stomach" to the prompt>cthulhu is summoned
>>101864591SAI is still acting as if they have the monopoly and that basic shit like lying a girl on grass should be celebrated, uhhh were they in a coma those last few days, we have flux now, they are done
>>101863765>>101863805Don't listen to this guy: if you want NSFW, use sdxl with any one of a thousand "pony" models. Flux cannot do NSFW at all reliably and is most likely intrinsically censored. Maybe not as bad as SD3, but it's not the tool to use if you want to generate anything involving women without clothes on. This may change eventually but it's not true now. Also, flux has no concept of artist style that can be invoked at all reliably either. It is really not very good except as a novelty. But novelty can be fun in its own right. Depends on what you're after.
>>101864591>hands out of frame
when will we get local Dark Rey feet gens bros?
>>101864802flux is good for shitposting and realism right now. also stylized text
How the fuck are people shitting out such good looking Flux LoRAs? Is it experience gained from SD or is Flux just that good to train on?
>>101864871is it difficult to train LoRas for Flux?I have a 4090 is that enough to do it?and how long does it take?
>>101864871If the base model is good it most likely already knows the concepts you are trying to teach it. You only need a little bit of training to bring it out. Even a textual embedding would probably give decent results for most styles.
>>101864864I love these slightly changed prompts
>>101864871if you're talking about celeb loras, its because flux already kinda knows the celeb so it doesn't need much to learn what to doartist style loras are still shit from what ive heard and seen
>>101864889A 4090 is apparently enough, this guy did it on a single 4090.https://civitai.com/models/638000/arnold-schwarzenegger-1990s-flux-lora
>>101863648>take a perfectly good base model>turn it into deep fried 1.5 slopkill yourself
>>101864927so 1500 steps for a LoRa?
>>101864938Sometimes I wonder if people's monitor color calibration settings are incorrect. I think most are, their gamma and contrast and all that shit is fucked up which betrays their eyes.
>>101864946Loss graph tells you how many steps
>>101864938this
>>101864917Doesn't really explain why it has no easily accessible concept of stuff like "Impressionism" to begin with. The loss of generic style terms seems like a pure regression.
>>101865102Try the german translation of those styles
>>101863805>>101863820>erotic>coveredThat doesn't sound too badI don't want photoreal or semi photoreal stuff like yhe ones posted in this thread, I like the 3dish stuff but not 3d stuff I got from darelites fantasy mix and anime style. Ir worked extremely well
>>101865102Simple, because the VLM that tagged all the images didn't know "impressionism" it just knew "a painting of"
>>101864999where can I read more about that?
>>101863829Base Flux? No, I don't believe you. catbox it right now
>>101865187I even used Kraut quotes around „Impressionismus”. It is a little better, yeah. Might be onto something kek.>>101865199Should have just used ChatGPT.
Flux makes some cool wallpapers but damn this shit takes long>picrel with 33 steps on a 4090 takes about 8 minutes
>>101865348>33 steps on a 4090 takes about 8 minutesnigga what are you doing
>>101865348User error
>>101865360what do you mean?>>101865384elaborate
>>101865389nigga a 4090 should generate that in ~30 seconds
20 steps doesn't give consistent quality, 30 seems to be the sweet spot
>>10186540430 steps kills diversity, 20 steps seems to be the sweet spot
what is the difference between /ldg/ and /sdg/
>>101865435one is for diffusion models talk, the other is for diffusion models created by StabilityAI talk
>>101865435/ldg/ is the based thread/sdg/ is the tranny containment thread
>>101865416yeah it's also true, less diversity but more consistency though, especially on text
>>101865416that's not good diversity, it is diverse because it's not converging enough and can have horrible lows, at least at 30+ steps you know the gen you're getting isn't a lucky gen, you know that if you touch other seeds you'll get the same quality
>>101865454all diversity is good, chud
>>1018653981920x1080 ? no way.it does work faster if I use Euler instead of Heun tho.
>>101865495>1920x1080 ? no waywayHeun does twice the steps, are you telling that that image took you 4 minutes?A 4090 should do it in 30 seconds. It does a 1024x1024 image in ~15 seconds
is it possible to train loras for FLUX with 16gb vram?
>>101865518fp8 takes 23gb for now. Currently grim for 16gb cards.
>>101865303It knows who van Gogh is, it just can't copy him.
>>101865517this one >>101865495 took about 3 and a half minutes.>A 4090 should do it in 30 seconds. might depend on the workflow tho, I noticed some workflows take longer for some reason.
>>101865435/sdg/ is for schizos and avatarniggers/ldg/ is for frens
>>101865542just post the catbox so we can point and laugh at stupid shit you did to make a 4090 take longer to generate an image than a 1060
>>101865533>fp8 takes 23gb for now.if you put the text encoder to the cpu it's only 12gb for the vram, and it's still fast
>>101865563https://files.catbox.moe/k0x9a7.png
>>101865538
>>101865554fr?
>>101865599I hate how 32GB ram is barely enough. Can't have anything else open.
>>101865599Pretty sure Simpletuner purges the T5 before training starts. "2024-08-12 00:08:20,290 [INFO] (__main__) After nuking text encoders from orbit, we freed 9.11 GB of VRAM. The real memories were the friends we trained a model on along the way."
>>101865653it's not barely enough, ComfyUI isn't using it optimally
>>101865653ram is cheap, buy some more my nigga, I'm at 56gb and I'm feeling good
>>101865605CFG with Flux Dev is a meme
>>101865710not a meme at all, you're delusionalhttps://imgsli.com/Mjg1Nzk5https://imgsli.com/Mjg1ODI5
>>101865710>>101865731so what did I do wrong??
>>101865745ok let me look at your workflow, did you add any flags into your .bat?
>>101865745how fast is it with CFG=1.0?
>>101865724its just a bunch of schizo stuff jumbled together to confuse the models into generating weirdness.https://files.catbox.moe/w840mz.pngwhere it says "painting of" you can make it a painting of whatever, not just a cabin.
>>101865745do you only have 1 gpu? if that's the case you can already remove the Force/Set CLIP and VAE device
>>101865768>its just a bunch of schizo stuff jumbled together to confuse the models into generating weirdness.damn, i thought you found a way to get a consistent art style with flux. oh well
>>101865745your image, the one using Heun, takes 10 minutes on my 4060, a 4090 should be close to four times faster
Why can't ComfyUI load the unet straight to the GPU? It loads fully into RAM first and only goes to the GPU when the sampler node runs so it's just there with T5 on the first gen slowing things down.
i keep getting this error message when running ImageSegmentation in ComfyUI. i realize it has something to do with TensorRT but i am unsure what that is exactly or how to fix it. Researching the problem led me to some reddit posts but they seem to be too old and just give me errors saying the versions im trying to install dont exist.
>>101865869I think that's just how computers work anon
need more like
I was expecting more of a silhouette, like seeing jesus in a tortillahow do I say that
>>101865869yeah I don't know either that's weird, if that can help you can use this script to force the model to only be on your gpu with OverrideMODELDevicehttps://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101865926holy shit that looks beautiful, how did you make that style anon?
>>101865915no, anon, it's not>>101865943I use it, works fine to pin T5 to the CPU but using it to move the unet to the GPU early creates issues with Lora loading and probably other memory management issues too since it's a little hacky
>>101865938first 20% of steps: black SVG icon of trump on yellow circle with orange outline, seen at anglerest of steps: vague likeness of trump visible in slightly charred pepperoni pizza
>>101865938ask bing or claude to make a boomer prompt so that the model understands better, or go for CFG 6 + GuidanceNeg, that helps aswell
>>101865877You on windows or linux
>>101865975linux
>>101865967>but using it to move the unet to the GPU early creates issues with Lora loading and probably other memory management issues too since it's a little hackyNow that you say that... I had no problem with loras with this Override stuff, but after I updated yesterday, everytime I load a Lora I OOM yeah
>>101865981What distro
>>101865988fedora
>>10186540425 steps seems really good, you get diversity and the text is good aswell
>>101865951it's lost to time, it's from 2 Halloweens ago
>>101865996Do you have mlocate installed? If not, install it, run sudo updatedb and run locate libnvinfer and see if anything shows up
mlocate
sudo updatedb
>>101866016>/.local/lib/python3.12/site-packages/tensorrt_libs/libnvinfer.so.10>/.local/lib/python3.12/site-packages/tensorrt_libs/libnvinfer_builder_resource.so.10.0.1>/.local/lib/python3.12/site-packages/tensorrt_libs/libnvinfer_plugin.so.10but this is not in the virtual environment. can i pip install these somehow?
Style keywords are "poisoned," which means you have to write everything out and hope you don't accidentally trip a different filter. This thing is going to need to be almost fully retrained to be of any use to anyone.
>>101866159>Style keywords are "poisoned," which means you have to write everything out and hope you don't accidentally trip a different filter.Such as?>This thing is going to need to be almost fully retrained to be of any use to anyone.I wish anon, I wish, but I doubt it'll happen...
>>101863852>When I try to load this lora I got a OOM error,>>101864264>bro same, keep getting OOM with using one GPU when adding the Lora.>>101864329>It's weird for me too. It will oom on one generation, I'll stop and try again and it won't oom on the next. I don't know why.>>101864343>After yesterday's update I get a guaranteed OOM on the first run, but all the following ones work just fine.Looks like we're not alone, Comfy fucked something uphttps://github.com/comfyanonymous/ComfyUI/issues/4338
>>101866190>Comfy fucked something upFUCK
Euler CFG++ is a fucking meme, it makes real people look like plastic: https://imgsli.com/Mjg3MTI3
>>101866174Well, afaict many major art movements (Baroque seems to work but a lot of them don't or don't well), every artist name, etc.. Like "an impressionist painting of flowers" produces something that is not a work of impressionism, so impressionism is functionally a dead token. So then what do you do? Describe the quality of the brush work to it? Or you just don't make a painting, I guess. The amount of stuff this thing doesn't do well for something of its size is fairly staggering lol. No wonder people are getting fairly good performance out of aggressive quants. Most of the layers are probably filled with functionally dead nodes.
>>101866280>Like "an impressionist painting of flowers" produces something that is not a work of impressionism, so impressionism is functionally a dead token. So then what do you do? Describe the quality of the brush work to it? Or you just don't make a painting, I guess.you can get the styles working if you go for CFG 6 + GuidanceNeg 10, have you tried it? >>101865731
what's the name for a thing women wear when they're in a convertible carI thought it was a wind scarf but I'm just getting scarves. I've tried babushka, headscarf and bonnet, but no dice
>>101866353believe it or not, 'convertible scarf'
1Asa-ing made|wasted my day again
Hey dynamic thresholding anon, why is it white? This is your same workflow I just tidied it up.
Have Flux devs said anything official about artists and styles? They're the only ones in capacity to tell us what is going here.
>>101866611That gen is really crispy
>>101866620They haven't said shit about fuck, I don't know what you're expecting. You know exactly what has happened. A VLM saw a painting by Greg Rutkowski and said "This is a painting"
>>101866620In particular, it seems counterintuitive for them to cripple their model on the one thing no one else is doing (especially closed competitors like MJ, they are just asking for someone to release a better model eventually and replace theirs). Hopefully a v2 or 1.2, etc... of Flux fixes the issue.
>>101866635Yeah, but we don't k ow if this was intentional and therefore won't be fixed (and thus the community is expected to fi etune back in all classical art styles, concept artists, etc...) or if it's an issue with how it was trained that they would work on.
>>101866643>Hopefully a v2 or 1.2, etc... of Flux fixes the issue.Won't happen, at this point it is us who should continue the pretraining/finetune
>>101865997i do 50 steps, unless i'm using heunpp2 then its 20-25
>>101866664It's obviously not "intentional" it's just a side effect of using a VLM to tag your images. The VLM would have to know the artists in the first place in order to tag them correctly but it won't. Take any artwork and put it into chatgpt and ask it to describe the image, that's how this works, with billions of images.Nobody manually looked at them each to make sure it was right.
>>101861802uuuuoooohhh androgynous demon tomboy erotic!
>>101865997>>101866668Hold on to your panties for the ultimate 1girl
>>101866620No and I doubt they will. We're probably more likely to get instructions from them on exactly how to prompt for hardcore pornography than we are to get info about what IP went in there lol.
>>101866668Still has that weird belly button
>>101866698holy kek
>>101866707it's a magsafe charging port
>>101866668nice clit piercing
>>101866684Yeah but there's an easy fix, you just give the VLM context like artists, mediums from the metadata and tell it that if there's any in there it keeps it. Obviously based on the nature data can be scraped this would be very doable. I refuse to believe they just took untagged images while SD can do a much better job from just random tagged images that appear on LAION.
>>101866719>I refuse to believe they just took untagged images while SD can do a much better job from just random tagged images that appear on LAION.I believe that, it's no coinscidence that Flux is so good at prompt understanding, LLM captioning has probably being used for easily 80% of the whole dataset training if you ask me
>>101866719How easy do you think that is, to give a VLM artist context? If you really believe this to be an easy fix please go ahead and train a VLM with this context, the entire imagen community will suck your dick.
>>101866719The model just happens to coincidentally suck at the two things that make imggen controversial and be really really good at everything else.
20 steps vs 500 stepshttps://imgsli.com/Mjg3MTM4
>>101866719>while SD can do a much better job from just random tagged images that appear on LAION.because those captions are shitty but often true regarding the names of the people in it or the artists that made it.
>>101866807>because those captions are shitty but often true regarding the names of the people in it or the artists that made it.this, at this point you give the laion captions to the VLLM to help it, it would make a fine combo
>>101866747"Here's a file with some metadata. You will caption this image, while paying close attention to the metadata. If it contains an artist name, including "by X". If you know the medium or style, include that as well. If not, include the source name E.G. artstation" that's literally it.
>>101866832It really nails reflections
>>101866846So you need to feed it shit tagged data which is the reason SD base models are so garbage?No.
>>101866862No, you're telling it to caption it as it normally does, but then analyze the metadata for artist names, styles and source and append them to the end in whatever format you desire.
>>101866862a smart enough VLM could work with it just fine
>>101866862Well, if the alternative is having to make a LoRA for literally every style of art, then ...
>>101866892>>101866900Are you sure the metadata is correct, clit eastwood?linkin_park_the_real_slim_shady_rubia_real_rare.mp3You are talking about billions of images. The only way to do this, is to tag every image manuallyORFine tune with LoRAs of these concepts, and since you only need about 20 images to train a LoRA this is a MUCH simpler solution.Then you leave it up to the highly autistic community.
>>101866935That's why you collect the images properly and make sure they have proper tags, and you also tell the VLM that if it's not sure just leave the info empty. But erroneous info would still perform much better than what we got.
>>101866935>Are you sure the metadata is correct, clit eastwood?I was just looking for a LAION browser to find that again but it's all dead now.Like I said, a smart enough VLM can work with even the shittiest captions. It just doesn't exist yet.loras as simpler but not a substitute for actual training.
>>101866960Please, by all means Anon, build this VLM. You will get 40 million is series A funding.
Is comfy always going through the entire image during inpainting, even if I masked a small part of it? I mean, it does change only the masked area, but it still renders the entire image, thus taking the same amount of time as genning the full picture.
>>101866955>That's why you collect the images properly and make sure they have proper tagsSo manually? Are you going to manually tag billions of images? There aren't enough literate Indians in the world for this task.
>>101866985shut the fuck upalways with the "if it's so easy just build it, you'll get rich" stupid ass bullshit when you run out of things to say
>>101866998Now do the painting in the style of Monet, and then Da Vinci, Picasso and Michaelangelo.
>>101867018Literally yes. It's the same shit when people complain about something in a video game, who have never coded a line in their life and they say "Oh it's so easy to fix! lazy faggot devs"
>>101866985You don't need a massive VLM to fix the issue we're having.You're vastly overestimating how hard it is to teach the model concepts of artists. There's a wikiart dataset published to huggingface. Just literally using that and giving it to the model would suffice. (Replacing any image that is duplicate). Most of the artists it needs to know are all congregated from a few sources. This is not rocket science, MJ isn't as bad as SD at prompt following yet you can ask it to give you a variety of styles, mediums, and artists no issue.
>>101867039because the idea of smarter VLMs is so outrageous, as if 4o doesn't existyou stupid mouth breathing motherfucker, shut the fuck up
How do I con comfyui into ignoring that .1 version difference "requirement" on OS install python without three pages of CLI and 8hr of arch wiki archaeology
>>101867021This is "Monet"
>>101866707flux is usually good at avoiding those. maybe it's related to me going crazy with the model shift values
>>1018670694o is dumber though.... Anon do you know what you're talking about?
>>101867082Maybe. Always good to put belly button in the negs.
>>101867089dumber than what, anon, DUMBER THAN WHAT>Anon do you know what you're talking about?YES, SO ANSWER THE ABOVE SO I CAN RIP YOU A NEW ASSHOLE YOU FUCKING IDIOT
>>101867082Belly buttons? They are easy, just say "her navel is showing"
I'm the debo
>>101867079"da Vinci"
>>101867109Dumber than yo momma
>>101867115now make her pregnantthen tell me your prompt because its hard to get the first trimester like this
>>101866862It's perfectly doable, man. You wouldn't even need to manually tag anything. There's no reason why these models shouldn't know every major artist and art style under the sun. It would be trivial to write a script to scrap wikiart and to automatically tag the images. Everything is already well tagged. You could even easily avoid anything that still isn't in the public domain, if that were an issue.
>>101867141Flux hasn't seen a single phallic shaped object in its short life.
I stopped worrying about finger errors and love them.Its fine if it works.
>>101867120
12b and the 1girls look like sd 1.5 gens. why?
>>101867141Anon she is not first trimester. She is well into the 2nd. Women only really begin to show in the 2nd trimester. Your pic is at least 6 months pregnant
>>101867180That's just cumbloating
>>101867177Because we are all promptlets figuring out the new way to do shit.
>>101867141https://files.catbox.moe/cm5e16.png
Steps testinghttps://imgsli.com/Mjg3MTQ4
>>101867197>Her pussy is really fat, there is a lot of fat in her pussy area. Anon your brain is something special
>>101867209where 25, 30, 35, 40 and 45?
If you prompt:> Art containing signature "[ARTIST]"that can retrieve some stuff
>>101867177They used synthetic data likely during the DPO/RLHF stage. Obviously SD 1.5 doesn't look aesthetic so it's a problem.
>>101867241I can do those too if you like after my current set.
>>101867251See, now this is a good fucking idea. The VLM would definitely be able to see that shit and it could mimic art styles from that. Post some examples, Anon.
>>101867209brap
yoga
It cannot differentiate stages of pregnancyLeft is 9 months pregnant, right is 1 month pregnant.Every month in between looked the same
whenever you prompt pregnancy it is never yours, you are a cuck
>>101867344I don't really care. I already have a son, my genes are safe so long as he doesn't become a faggot.
Should I trade in a 7900XT for an A4500?>3090The end goal is to stack them inside of this shit box. Can't stack 2 3090s
>>101867209can we get a catbox for this one too?
>>101867372If your intent is AI workloads then yes.
so what happened to the "clip_l understands artists" thing? in my testing, putting a photographer in the clip_l box certainly does work.
>>101867384Yes, will deliver with the 25-45 steps test
>>101867209>>101867241https://imgsli.com/Mjg3MTUx>>101867384https://files.catbox.moe/zy3mb2.png
>>101865753what flags?>>101865758about the same>>101865771>do you only have 1 gpu?yes>if that's the case you can already remove the Force/Set CLIP and VAE devicecan you post a better workflow then?I have no clue how to put this spagetti shit together.>>101865810>a 4090 should be close to four times fasterso why isnt it?
>>101867390Certainly does, or certainly does not?
>>101867279> painting containing the signature look, style, and feel of "MONET"Not like perfect but the resemblance is there
>>101867400based, thanks for catbox anon!
>>101867318I'm 10000% sure we're getting the belly slider lora eventually, along with the overall fat, breast, age sliders like it was with sdxl/pony
>>101867386It's primarily an AI workstation. What's nvidia like on a linux system?
>>101867412it works. try "a person" in t5, nothing in clip_l > baseline fluxslop with the occasional weird outburst. now add a photographer with a distinct style into the clip_l, like "paolo roversi". voila. or try nobuyoshi araki. this is without negative and a positive guidance around 2.5
>>101867473https://imgsli.com/Mjg3MTYw/2/1It does not>>101867431Will test this next
>>101867409>about the samethat should absolutely not be the case, CFG at 1.0 means half the work, half the timecan you try using the first workflow from here? https://comfyanonymous.github.io/ComfyUI_examples/flux/just run it and see the speed
What is the actual prompt token limit? I've been putting in 1000+ and getting decent results, but it seems like it forgets some stuff from earlier in the prompt.
>>101867496Oh by the way the artist I'm trying to prompt is inflation4furs.
>>101867431>painting containing the signature look, style, and feel of "MONET"I was really hoping this would workhttps://imgsli.com/Mjg3MTYx
>>101867496Wow, bro. These bitches are hot. Now I get why this Picasso guy is so famous, bro.
>>101867504Technically 512 but some anon the other day said it it's actually only 256.Could be because we are all running fp8?
>>101867431Picasso.
>>101867550When I am doing tests like this I want something good to look at with lots of little details. That's why I like using this prompt for it. Nice ass, lots of intricate patterns.
>>101867566this thing was not (or barely) trained on dead artists work; that we can conclude
>>101867566Also Picasso.
>>101867575Could eat those asses for days, bro. I really hope soon they make 3D printers that give us waifu robots, bro. Would sell my car to get one.
>>101867431This works if not prompting anything specific it seems. Maybe my prompt was too complex. >painting containing the signature look, style, and feel of "Greg Rutkowski"
>>101867564I heard that 256 was for schnell and 512 for dev. I read somewhere that t5 doesnt technically have a limit but it gets more retarded when you go above 512.
how the fuck do you have 512 token prompts, why are you writing paragraphs
>>101867512>painting containing the signature look, style, and feel of "inflation4furs"
>>101867593The photorealism bias hammers impressionists and abstract artists, I think.
>>101867620I'm giving my prompts to an LLM and ask it specifically to expand them to a 500 words essay. Works like a charm. Unironically.
>>101867628catbox one so we can laugh at it not using even 1/5 of what is in the prompt
>>101867566>>101867596>12B parameters>the dataset doesn't contain PicassoHow do you fail something so simple?
>>101867646It likely does but was not captioned with "by Picasso" by the VLM they used.
>>101867658Oh, right. Yeah. Good call. They probably didn't even think of it. They were so excited by the VLM that they overlooked this. It should be easy to fix though.
>>101867658What VLM did they use? If they told us just that we could go back and unfuck things on our own lol.
>>101867703I'm calling it VLM but could be a caption model.All we know is that whatever they used is dumber than what OpenAI used for DALL-E 3, so it could be anything openly available VLM/caption model.
>>101867618i really like these, what about inpainting the keyboard and screen for more detail?
>>101867680It is easy to fix. Look at the LoRAs of celebs that were trained with only 25 images. It already knows them, it just needs to be told who they are.Which is why for Flux, embeddings is probably a better way to go, rather than LoRAs. It already knows all this shit, it just needs to be told "this face is Emma Watson"We could get packs of tiny ass embeddings that bring out all the people and styles we want.
Ready to go with the next bread...>>101867704>>101867704>>101867704
>>101867724i cannot do that
>>101867729>We could get packs of tiny ass embeddings that bring out all the people and styles we want.But no one does embeddings anymore.Is there an embedding trainer for T5 yet? Would you train just for CLIP, both?I wish they had used just T5.
>>101867719Salesforce has an open source BLIP-T5 caption writer. Wonder if it was something like that.
>>101867754Clip is there for a reason, I just don't know what that reason is. Probably so trainers can re-use their data-sets
>>101867500I tried it again after restarting my browser and now it only took 3:25 minutes with the old workflow.on that other workflow with same seed, sampler and steps it took 3:02 minutes.so CFG to 6 or 1 is only a difference of 23 seconds?
>>101867813something is seriously fucked with your setup, the example workflow should take under 20 seconds, are you sure you have a 4090?
>>101867813and no, CFG at 1.0 should half the generation time, it is literally doing half the work
>>101867318try "early pregnancy"
>>101867832>are you sure you have a 4090?yes
>>101867898how are the temps?
>>101867913looks normal
>>101867929have you ran benchmarks/games before on it? how was the performance.
>>101867985>have you ran benchmarks/games before on it?yes>how was the performance.as expected from a 4090
>>101868081well, only thing left to try is alternative interfaces like A1111 or Forge to see if the issue is with Comfy.
>>101868095ok will do.
Any disadvantages running models on mac comparing to graphics card? (excluding not liking platform)
>>101866609update ComfyUi anon >>101867409>can you post a better workflow then?bruh you just click on those nodes that have "Force/Set" and click on delete, how hard is that?
