Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101967932>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>GPU performancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>Maintain thread qualityhttps://rentry.org/debo>Related boards>>>/g/sdg>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>mfw
we are in the new age of memes and image creation. and to think after sd3 people thought open source gens would plateau.
Shilling my custom script for joy captioner again with batch processing and easy trigger word insertion.https://pastebin.com/p2PWgJZK
>>101969649based miku, goth booba & of course tasty BMW enjoyer.>>101969678the time is now.>>101969672I thought more of you.
>>101969686put that in a github anon
>>101969649I haven't been in these threads in a long bit but HOLY SHIT these still look really bad. This is what AI bros have been hyping up?? You He's been doing all this time??? LMAO
>>101969686why are you posting kids here
>>101969691>I thought more of you.what does this mean
a gen of pepe genning mikufluxception
>>101969699Local imagegen is, and forever will be, full of ugly plastic slop.
>>101969697But it's just stolen code from that huggingface repo and gpt slop. I only made it for a very specific task and figured people here probably want it too.
>>101969716Pepe is sitting at a computer in an office with a CRT monitor. He is smiling and typing on his keyboard. On the monitor is Miku Hatsune who is smiling. Pepe says "feels good, man" in a white speech bubble. On the wall is a painting of Pepe in art nouveau style.enjoy this SUPER RARE miku pepe.
>>101969736>But it's just stolen code from that huggingface repo and gpt slopyes? it's called forking a repo, and there's nothing wrong with that
>>101969715what makes you snipe these threads to dump your >mfw nonsense? we could all be one gang. "laser focused", remember? that was me>>101969720then why don't you just fuck off>>101969731YES lol
>>101969720>imagegen3 letter agent glow so bright
>>101969752He does it to piss you off. His existence revolves around trolling, nothing else. Just follow the rentry.
>>101969776who is she
>>101969752>schizo on day 501 of stalking and harassing me*sleep*>I participate in a thread"omg how fucking dare you"
OK training a lora on 2 subjects in the same pics, let's see if this works or just averages their faces.
>>101969795He's so butthurt from using the same copes for over a year straight, it's pretty incredible.
fp8/e4m3 weight model or Q4 model in terms of quality? or are they similar?
>>101969798Why don't you simply stop feeding them? Participate, don't react.
>>101969832Q4 is on par with, if not better than fp8
>>101969837When he doesn't get enough attention he starts drama for the sake of it.
>>101969818It probably works with the right LoRa DoRa whatever thing, settings and training data.Which these are is a perpetual mystery and barely any tooling helps to approach the good choices (IDK, with A/B/C comparisons).
>>101969820>>101969798>>101969741>>101969716>>101969699>>101969686>>101969672Real talk, why is this general full of pedophiles?
>>101969862I'm just using the exact settings I used for my single subjects, which worked perfectly. Will see how it looks in the end. I know people have done it before on SD, I have to assume it works on Flux too, but who knows for sure until someone does it eh.
cartoon illustration lora:
>>101969890She has SD hands. Is that just this gen or is it the LoRA?
>>101969837he seemed like he was engaging with me earnestly for a moment and I got baited. I'm at about a 95% anti-feeding rate these days>>101969832>>101969847someone posted an interesting quant comparison earlier today. iirc q4 was still below fp8 but its debatable>>101969890>cartoon illustrationis that not just anime?
>>101969900some outputs have a distinct style, the disney lora is clearer difference tho
>>101969900>q4 was still below fp8 but its debatableyeah very subjective. The main selling point is that it's less expensive for still excellent quality. Either way I am perma using Q8 so I don't really know.
>>101969865Nobody likes you.
how do you get harrassed on an anonymous image board?
>>101969922By telling everyone who you are
>>101969921>>101969910is the extreme blur inherent to flux or its the prompters here that are really shit?
>>101969922Making yourself identifiable and trying to become a character
>>101969686based I also did my custom shit to enhance the caption by injecting some external context, I'll be using to train my loli loras since it is excelling at lolis
>>101969956sick, prompt?
>>101969956nice
>>101969736that's GitHub in a nutshell anon
>>101969964old man playing with his balls
>>101969922be debo, pw, cumfy, trani, any avatar faggot
>>101969964some llmslop trash
Answer me this: if one were to train lora using 10 perfect looking AI images which can't even be told apart from real art, is it still worse than training using 10 images made by humans that are lower quality and objectively shittier looking, than the AI images?
>>101969994better version
>>101969882Yea, I suppose you can only try.
Flux video when?
>>101970007Of course not. However ensuring a good selection of these images is nontrivial either way, neither aesthetics NOR most aspects of technical image quality can be really easily and reliably sorted out exactly. For some trainings the remaining uncertainty isn't a large deal, for others it might be the issue.
>>101969927The biggest culprit is screwing around with the FluxGuidance node default value of 3.5 and going over or under a range of 1 from that value aka a value outside 2.5 to 4.5 without dialing to a specific value that works. For me, I like 2.1-2.2 which is outside that value but works for me with realistic generations.Sometimes it's a sampler and scheduler issue. As in the schedulers and models don't produce anything good no matter what you throw at it either because the math doesn't work or was hardcoded for SD only. I made a post half a day ago with a comparison image I found.>>101961683Finally, I haven't had a good result with negative prompts at all or a CFG above 1.0 for realistic images that isn't blurred, changing the guidance helps but it doesn't solve the issue.>>101970007If this differentiation can only be done on the visual aspect front, I would still say no because who knows how the diffusion/image generation process encodes information in an AI image vs a real one and what an AI learning from those images would take from those images having that data. If that association and data didn't exist or was identical to real life images, then yes, but we're a long way off from that. LLMs are the clearest example of that with text, where people see just so much slop writing because of AI ingesting shit content based on other modeels to train based on other poorly trained models. That's why Command-R got the praise it did in how it was unslopped vs the rest of the models releasing.
>>101970007I think there is objective research that says repeatedly feeding AI images back into AI training causes a feedback loop that makes the image look way worse than using authentic data.
>>101970007you might not be able to at a glance, but baked into those 'perfect' AI images are the AI's biases. maybe it prefers a certain color grading, a certain shadowing, a certain angle or framing, etc. when you train AI on synthetic data, those biases get reinforced. this can reduce output variety or cause overfitting
>>101970063>>101970007MODEL COLLAPSE
>>101970059This is not proven for different AI or with a selection procedure in between or anything.In fact with a perfect selection it'd get *better* until it's all perfect just because of the introduction of noise in these processes.
pc98 lora migu
>>101970076very nice. three thumbs up!
>>101970088you're welcome!
Not my problem.
>>101968256okay got Q4 Guff and got T5xxl_fp8 to actually show up in forge but the vae is not there, i put the ae vae on the vae folder but it doesn't show up on the UI, is there another vae that i need to use? pic related is how it looks.is this okay?
cheers from bar miku>>101970161vae in vae folder and it should show up, try refreshing the gui
>>101970068>MODEL COLLAPSEfun prompthttps://suno.com/song/67a0380f-fb1f-4c41-aaec-3a6f4053a377
>>101970161also make sure it's ae.safetensors in vae folder and 327,446kb
we dont need windshields, it's the future
>>101969720Nah, those who post plastic gens just suck at imagegen. There are already Flux realism LoRAs that look as realistic as MJ with good prompts.
>>101970174the one i found on huggingface is 335mb, is it the same thing?
>>101970197that's the spirit!
reminder that if you're out of ideas it's fine to just not post anything
>>101970213thats it, just put it in the vae folder, if it wont show up try restarting the guinow im not 100% sure if thats the vae for schnell model only, they should all use the same one though. also you need clip L to show up along with the vae for it to work (2 clip models, 1 vae).
>>101970198For instance pic rel is from a realism LoRA
>>101970228okay i think i got it, s this how it should look?
>>101970252waiting for the sovl lora
>>101970228This reminds me, someone published a more accurate Clip L:https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/blob/main/ViT-L-14-BEST-smooth-GmP-ft.safetensors
>>101970261whats the c9 file? isnt it supposed to be t5xxl_fp8_e4m3fn.safetensors (or fp16)or maybe thats a new encoder I dont know about yet
>>101970268
>>101970273nah, i think i just misclicked, there, i think i got it now.
>>101970283looks good
wtf is flux
>>101970303current overall most popular imagegen model type
>>101970303new SOTA model with capabilities on par with dalle but has dog shit aesthetics. waiting for good finetunes.
>>101970313>dog shit aestheticsmore like mediocre than dog shit, we've seen worse
>>101970303New hotness, large smart model with amazing prompt coherence but an incredibly slopped aesthetic and poor knowledge of art styles and artists.Yet to be seen if the last part can be corrected.
>>101970356and now with a ps1 style lora
>>101970346>Yet to be seen if the last part can be corrected.Demonstrably has been fixed with LoRAs, finetuning almost certainly can too.
>>101970346It's not like the Flux technology won't be able to understand. It's more of a question if community trainings can achieve it soon.
Is there a node to load multiple loras at once in Comfy?
According to the BFL podcast their video model is>16x faster than SVD>Can generate much longer videos>Can keep object and character consistency not only in frame but between cuts
>>101970389All the art style loras I've tested so far also make the model much dumber in the process of teaching it the style, there's instantly a big coherence drop
>>101970397>BFL podcastlink?
>>101970397No way they're gonna drop the weights if all that is true
>>101970420>>101970426https://www.everand.com/listen/podcast/759758667
>>101970388and pixar:
>>101970389nta but having to switch in a different lora for every individual style or artist is gonna be a pain in the ass
>>101970465Why? You can just write the LoRA you want to use in your prompt. You don't have to switch anything.
>>101970494No you don't get it, the model must know anything and everything.
i just want a model with the capabilities of flux and the sovl of base pixart, man.
Going back into genning after 1year away. Thinking about giving SDXL a try.Also, is there anyone here competent enough to make a LORA or something outta this style? This guy's works seems like AI, surprised if theres not a LORA that replicates this style yet.https://x.com/KGUY1920
>>101970525>Thinking about giving SDXL a try.SDXL? That's a baby's toy.
"realistic anime" lora
>tfw can't post any of my gens today because they are all LoRAs of my actual face and I will dox myself because of how exactly perfect it isI'm never making it into a collage again.
>>101970537What's the new "meta" for AI art? I am messing around with the good ol' SD1.5 before hitting the bed and diving in Controlnet again and all that stuff tomorrow.
>>101970514>i just want a model with the capabilities of flux and the sovl of base pixart, man.2 more weeks
>>101970584Flux. Anything and everything Flux. SD is a cave painting compared.
I improved the Tonemap node for those doing some CFGmaxxing and using this node : https://files.catbox.moe/s5js6c.pnghttps://reddit.com/r/StableDiffusion/comments/1evvjcq/adding_new_features_to_the_tonemap_node/>Hatsune miku skateboarding, her speech bubble says: "I'm loving it!", 50's comic book stylehttps://imgsli.com/Mjg5MDA3
>>101970539last one, a disney miku
>>101970642Weird I made this same style of image, but different. Can I have your prompt?
>>101970642pixshartists i kneel
>>101970436>They are working on watermarking tech for all of our images and videosThat bit was obvious but it's concerning nonetheless. I tested one of the AI detection tools and basically all of SD is watermarked regardless of the LoRA you use. Apparently the only way around it is to train your own VAE or alter it in some way.
>>101970645I just checked the OP, but can you give me a general direction to what Flux is? and can I run it with 16gb RAM and RTX 3050? Sorry for being annoying, last question.
>>101970659I genned these way back early pixart days, prompt was simply "Minimalist aesthetic anime art of a girl smoking"
>>101970388That looks better than the SD PS1 lora's I've seen, you using this one?
>>101970700>I tested one of the AI detection tools and basically all of SD is watermarked regardless of the LoRA you use.That's not necessarily watermarking
>>101970293>>101970168I got this error on the console after waiting for like 10 mintues, any tips? the model (Q4.guff) was inside a .rar and i extracted it, should i have done that?
>>101970735Aah yes, you were the one I stole the idea from in the first place.
>>101970525>sdxllmfao what year is it? 2002?
>>101970748I think that's it, looks familiar>>101970787yeah you need the q4 file, try updating forge with the exe and try again, gguf support wasnt day 1 I don't think. Or try comfy, works for sure.
Is there any Comfy Node that allows me to do the VAE render after X number of iterations?
>>101970388>ps1That's looks a lot more like a Dreamcast game than a PS1 one
>>101970814it varies, it is consistent with the low poly look overall though
>>101970804I don't understand what you mean. You mean pull the incomplete latent out and decode it? You will just get a blurry mess. You can stop the gen at step x, and pull the latent into another model and finish it there. Is it something like that?
>>101970814>>101970821PS1 has such a unique look with the geometry and texture warping it's impossible to capture with a lora I think
>>101970700Why is watermarking concerning?
>>101970794Cool to see I inspired someone kek. I still found the exact prompt and consistency with cigarettes works best in Pixart despite its small size.
>>101970795I stopped genning for 1 year and a couple months... can't believe anons are really shitting on SD, sad to see they fell so hard.Anyways, gonna sleep now. Unfortunately anon didn't spoonfeed me about Flux.
>>101970855>Anyways, gonna sleep now. Unfortunately anon didn't spoonfeed me about Flux.https://www.youtube.com/watch?v=stOiAuyVnyQ
>>101970841SD always struggled with cigarettes. Flux is better but pixart definitely wins the durry war
>>101970840Because then they can trace his pizza back to him.
>>101970827>I don't understand what you mean. You mean pull the incomplete latent out and decode it?Yeah, this. I think there's some value in it because it gives a rough idea of what it's going to render if you have many iterations set. So being able to see what it's doing every 10 iterations can be handy.
>>101970798>try updating forge with the exe and try againi did git pull before starting doing all of this, is that not it?>Q4 filethat's not the model?
>>101970840It's an issue with marketing. You can easily get banned or shadowbanned anywhere you post AI images if this watermarking tool becomes standard. Forced identification of AI images. I know IG and TikTok are already doing this.
>>101970889forge has an update file in the root dir, update.batq4_0 is fine
I don't get deformities on flux very often but when I do it's fucking funny
>>101970835Forget about the warping, the textures in the image are all blurry and filtered, which is pretty much anti-ps1 in aesthetics
>>101970885You can just turn on image preview to see it while it gens.
>>101970904total recall had a 4 boob woman
>>101970897>this watermarking toolWhich watermarking tool?
>>101970919No she had 3.
>>101970931ah, well the gen is perfect then
>>101970922The one the flux devs said they are working on. They probably already have one for Flux and they will likely release it with the video model.
>>101970942Most watermarks are pretty easy to bypass. Create multiple derivatives of your output containing subtle noise and posterization, combine and average, chances are you killed the watermark.
>>101970972Literally just upscale it in a non-watermarked model at very low noise.
>>101970984Or that yeah kek
>>101970855That anon is full of shit. SD is still king for being the most versatile and fully uncensored.Flux is great, but still limited.
>>101970989Only if you're talking about finetunes, and at that point it's moot, because Flux finetunes will rape the anus of any SD finetune. We're still working with a base model, never forget this.
>>101971005Flux is made for jeets and normies to shit out stupid AI slop for lowtier content. Besides SD any person can make their own Loras and stuff by themselves with midtier hardware.If you think Flux will be best at making anime/manga girls or NSFW you are delusional.
>>101970397I believe everything becomes faster and consistent the more it runs Text_to_3D and 3D engine internally.
>>101971027Go back to figuring out how to make a woman lay on the grass Lykon
>>101971027What led you to this conclusion? It's just the SOTA local model right now, people will tune and LoRa it.If nothing better yet shows up, it'll be the best at anime/manga NSFW at some point, yes.
>>101971051Only @ me again when Flux can make something like this
A blurry mess with shit loads of inpainting?Sorry, it just can't.
>>101971051Holy overbaked batman
>>101971073
>>101971083Much better. Looks good.
>>101970646
>>101970972No, I forgot to mention this, but with Hive moderation in particular, I tried with completely cleared metadata:cropping, resizing, rotating, taking screenshot of the image, taking photo of the monitor with my smartphone, jpeg compression, GIMP and Photoshop filters, steganography, neither of these work, but what worked for me:lots of noise (where image became unrecognizable), changing color/adding a bunch of contrast, works but image is now filtered and very bad, turn photo into drawing with filter: somewhat works, transfer histogram from working image to this, somewhat works but hive in particular is still going to be a painAll of the above only apply to SD, Pixart, Hunyuan, etc... even closed models.Flux: passes detection in the meantime. They are not doing anything special are these are images that are impossible to detect with the human eye, let alone zooming in. They are not training a model to detect images with errors. It appears they figured out the underlying sequences of VAEs or distributions of noise of SD, MJ, GAN, SDXL, Dalle, etc.. which is how they are able to detect so accurately AI images.
>>101971092Nice!
>>101969053SAUCE PLSSS
>>101971127What a dumb house. Their kitchen and dining room is completely open to the snow with no way to close it off.
>>1019711311girl, goth
1girl, goth
>>101971100>taking photo of the monitor with my smartphoneNo fucking way is the data loss in doing so insufficient to pick up on the fine differences in noise between seeds
>>101971027>Flux will be best at making anime/manga girls or NSFW you are delusionalDo you not remember how shit SD was at the very beginning? We couldn't make anything near SDXL and Pony levels. In a couple of years Flux will be great for porn too
>>101971147He's 100% full of shit, Anon.
>>101970646Holy fuck, Filmic + GuidanceNeg 10 is insanehttps://imgsli.com/Mjg5MDIx
>>101971155I just dumped a few Stable Diffusion images into Hive and all were predicted non-AI, and I didn't even modify the images other than cropping a 256x256 section.
>>101971150Let's go back in time for a bit and reminisce about how uncensored SD 1.5 base model really is. https://files.catbox.moe/ol6l07.jpg
>>101971100>>101971147>>101971155kek, got my first positive. A false positive, using a texture from the game Marathon 2: Durandal.
>>101971150>In a couple of yearsTry months, if not weeks, Anon. Have you seen some of the LoRAs coming out?https://civitai.com/models/649472/anal-art-for-flux
>>101971165Not even 1% AI? Predicted non-AI but still having anything above 3% is bad, I found they are almost 100% accurate on this too.
>>101971168a lack of base model censorship we will never see again, unfortunately
>>101971192They were all under 50% and said they were unlikely to be AI. The closest was 48%, but it guessed Midjourney when it was in fact SD 1.4. Whatever algorithm it's using, it's obviously not picking up on seed-specific noise patterns, or at least not very well.
Don't most ai image share a similar kind of noise? Why would prediction software not look primarily for that?
>>101971204Yeah but I got images like that too. But then drop a real image in there, and you will see nothing that high in percentage. Even below 50% has a high probability to be AI generated from what I've seen, for instance 40% which is the best I got in my case.
>>101971187Trained a LoRA on myself. The image is 100% AI. I feel like it just makes shit up.
>>101971192>>101971204Ok I just got two correct positives in a row with 99% confidence and Stable Diffusion identified, but both were using 512x512 images instead of 256x256. Maybe the alleged noise algorithm needs higher resolution to work properly.
>>101971220Forgot pic>captcha: g0yrd
>>101970917How do you get from latent to preview image though?The sampler will only output to latent once all steps are completed, no?Is there some trick I'm not aware of?>tricks are something whores do for money
>>101971242Comfy Manager, which should just be a pre-installed addon at this point.
>>101971188Yes, but every lora is very limited to a very specific pose and angle. You still can't make specific stuff like with PonyXL
>>101971242the latent is being processed by the sampler, it has access to the latent at every step
>>101971242>>101971251You can also add an arg to the .bat to show it without Comfy Manager but I don't know what that arg is because I have Comfy Manager.
>>101971235Here's a real image. I don't think it's making it up, the stuff it's counting in the percentage are whatever it thinks are AI.
>>101971252Pony is a finetune though so it doesn't matter. Once we get a ponyesque finetune for flux the coomers will all move to it.
Why does it normally take 1.5s/it then randomly take 2 minutes per?Every few gens I need to flush the VRAM
>>101971297buy more VRAM
>>101971273I'm the other anon and doing more tests, it does seem like it's getting the bulk correct now. Struggles with flat outputs like textures though, guessing that it's picking up on perspective/accidental non-Euclidean geometry rather than seed patterns.
>>101971281>Once we get a ponyesque finetunehopefully neverit's insane that redditors fucking praise him now for shitting out that sepia-ridden piece of trash
>>101971310If I had known flux would be so horny for VRAM I would have but my retarded ass thought a 4080 would be enough last year.
>>101971318It will happen whether you want it to or not, Anon. Porn has and always will drive technology.
>>10197131916GB? You're doing something wrong then, have you pinned T5 to the CPU yet?
>>101971329No, I have an i5, worth doing?
Just so you know, cooming several times a day and sleeping poorly, especially at our age, boosts oxidative stress and weakens immune response.imagegen will be the end of me
>>101971319>mfw saved money by keeping my good old 1070 and will be buying a new GPU just for Flux when it can reliably gen porn as well as PonyXL
>>101971328I do want a anime porn fine tune to happen but not from this fucking faggot.
>>101971335I have generated tens of thousands of images and seen thousands more from others.Haven't masturbated to a single one of them.
>>101970984>>101971100It detects the chins, you FOOLS. It's the CHINS!!
>>101971317I don't think there's a way for them to figure it out if it's artwork using simple lines, plus those you can just preprocess them in Photoshop or Illustrator, so at least those are safe for now.
>>101971334an i5 750?? Jesus, dude, that's 15 years old!
>>101970984>non-watermarked model at very low noise.All upscale models use SD. All SD models can be detected by Hive.
>>101971359It managed to get picrel with 99% confidence, which is far from the sloppiest 1girl I've seen
>>101971187Duh. A meta-stable (possibly Leela) AI made that texture. Bungie has been doing soft disclosure for years.
>>101971281>>101971318Apparently there will no be finetune pony flux due to license, pony have said they may finetune for Auraflow though.Things may change, who knows?
>>101971273Does it know that this is AI?
>>101971381kek
FUCKING FINALLY, Deal or no Deal models!https://civitai.com/models/662154/deal-or-no-deal-models-flux?modelVersionId=741008
>>1019713869.6%. Despite what it says I think it does know.
>>101971376it probably have seen plenty of novelai outputs at this point and got trained accordinglyused to be 0% back when nai was released
>>101971399Like it says one thing but it's actually thinking something else? Like a woman?
>>101971376Well, like I said Flux, perhaps other models like NAI are still not detected. It'll always be a cat and mouse game, once they are new models or optimizations will come out that are not.
>>101971334I have a i5 13500 and it takes like 3ish seconds to load and complete T5 on the cpu (fed by a 7GB/s m.2)So, yeah, might be depending on the gen of the i5
>>101970708Flex is a model in two variants. Dev (https://huggingface.co/city96/FLUX.1-dev-gguf) and Schnell (https://huggingface.co/city96/FLUX.1-schnell-gguf). The latter is a distillation (is that the term?) of the first that requires only 4 steps to produce decent results.The GGUF go in ComfyUI/models/unet/If you don't have t5xxl_fp16.safetensors or clip_l.safetensors already in your ComfyUI/models/clip/ directory you can find them on: https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main. You can use t5xxl_fp8_e4m3fn.safetensors instead for lower memory usage but the fp16 one is recommended if you have more than 32GB ram.The VAE can be found here: https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/ae.safetensors and should go in your ComfyUI/models/vae/ folder.You can use this node to load the GGUF: https://github.com/city96/ComfyUI-GGUFMore knowledgeable anons, please correct the errors in this post. Thanks.
I imagine in the future the text for AI will get even better to the point of being able to write whole page paragraphs into an image.
>>101971432>FlexIt's Flux
>>101971242>>101971251>>101971262Thanks, the argument is --preview-method "auto" | "latent2rgb" | "taesd"
>>101971002This one is almost perfect.
>>101971395wow, the model we all needed
>>101971442It's so weird to see the website layout make sense and then instinctively try to read the text to not understand anything
>>101971452Thanks
>>101971463Yeah like when I try to read in a lucid dream.
Yesterday I was pointed to a flux model that I was told should run on an 8GB card. Do I just stick it in the models folder in forge and run it or do I have to do lots of extra stuff like this guy>>101971432Says
>>101971100I was actually messing with hive recently too and managed to trick it. I made a 1girl with a white background in ponyXL and hive was 99.7% sure it was AI generated. Then I just selected all the white bg and replaced it with solid white in an image editor and it plummeted to 17% (probably slight noise fluctuations in the bg before). Next I edited various little color gradients and shading to make them solid color and it went to 12%. Finally just replaced the white bg with a color gradient and hive became certain it's NOT AI with a 0% chance. Pic related.Seems like most of its AI detection relies on picking up noise patterns in solid colors or smooth gradients.
>>101971532>Pic related.ANON, THE PIC, IT GOT LOST IN TRANSIT
>>101971349I envy you. I can't help myself and end up having painful dry cooms almost every day. I'm sleepy all the time now.
>>101971362No, a 10400k>>101971424Can't hurt to try, I will give it a go
>>101971424Really? 3 seconds to load and process? I have a 7700, T5 is always loaded in RAM and it takes about 6 seconds on a short prompt for me.
>>101971538what the fuck it's like I'm shadowbanned from posting images, it just won't work
>>1019715544chan is detecting your images as AI
>>101971444lmao
>>101971168SD adding a single artist fixes the issuehttps://files.catbox.moe/09wxgq.jpg
>>101971569I can gen asses in flux.
>>101971569Flux can do butts
>>101971432Also, you load the text encoders using a dual node.See example workflows: https://github.com/comfyanonymous/ComfyUI_examples/blob/master/flux/README.md
>>101971562That was fucking bizarre, everything works fine except that one image is somehow cursed even if I try to upload a screenshot of it. Here's the one before I put a gradient in the bg.
>>101971568wholesome
>>101971553yeah, it's very quick comparitively 4s max , I'm unable to get an actual timing atm, ill turn on the timer feature on comfyui later and post it somewhere.
>>101971553oh also its DDR5 6400
>>101971424>>101971541>>101971553Just tried, every time I change my prompt it takes 65 seconds. I am using fp16 t5xxl though. Will trp fp8
>>101971475As you get better you will eventually be able to read (and write), even in learned languages other than your native one. Meditation and visualization exercises help a lot. I technically have aphantasia (I just "imagine and pretend" that I'm visualizing), so that's no excuse. Your brain gets the training all the same.
>>101971568MOM I POSTED IT AGAIN
Can Flux reach Midjourney-v6-niji's level with a top tier finetune?
Is there a comfy node to just not load the negative clip but have a connection into the ksampler?Since it's unused in Flux it seems dumb to load it every time.
>>101971641>Since it's unused in Fluxit's not unused if you want to go to the CFG > 1 territory
>>101971646I am aware but I am not doing that so I just want to not load it.
>>101971610If the entire process is 65s then that's not terrible, my 1024 fluxd gen is 49s with fp16t5 4060ti 16
>>101971650you could download the comfy's workflow, it doesn't have a negative prompt at all
>>101971641There is a placeholder empty prompt node, but I forgot what it's called.
>>101971532Yes I know about removing background too, at which point it had dropped to 40% for me. The idea is to simply use color mapping, and/or a healing tool or a specific type of barely visible brush afterwards to paint over the entire image. Thanks for input.
>>101971658Not the entire process, the time to load the clips. Then it gens at normal speed after that. >>101971664Which one is that?
>>101971636Probably.>>101971628Damn...looks like a picture. Very comfy.
>>101971636It already is there, but anons are sleeping on the specific lora-https://civitai.com/models/647940/flux-atilessence-lora-test?modelVersionId=724910
>>101971636>ai slopyour example images don't look impressive at all, flux could defienetly beat that with a good finetune
>>101971641Just use the same positive prompt but keep CFG at 1.0. it won't be used, you'll get the exact same gens.
>>101971586But yeah I can imagine that being a thing in a lot of placing, auto stopping you upload images because it's AI
>>101971720this, the finetune theoretically comes out in 10 business days
>>101971718Note you have to be good at prompting it, kino doesn't just happen on its own-https://civitai.com/images/24856741https://civitai.com/images/24494747https://civitai.com/images/24456923https://civitai.com/images/24856807
>>101971616intersestnig
>>101971002Introducing the revolutionary NEDIU keyboard layout! Designed with the utmost ergonomic precision, the NEDIU layout is perfect for those who prefer to type with one hand while sipping their coffee with the other. This innovative design ensures that all the most frequently used keys are conveniently clustered together, reducing finger travel and increasing typing speed by 300%. Say goodbye to the outdated QWERTY and Dvorak layouts, and embrace the future of one-handed typing with NEDIU – because who needs two hands when you can have one super-efficient one?
>>101971720if this is ai slop, then everything is ai slop
>>101971725in 2 weeks, trust ze plan
>>101971740not prompt issue but that moron used schnell model
>>101971755you're crazy if you don't think those images don't look ai generated. don't get me wrong, they look good... for an ai image.
>>101971809you look good... for a frog
>>101971809I never said they don't look AI, but they just look good, and I wish Flux could reach that level, but for that I think it needs to understand more concept and stop always go for the most generic one
>>101971718Where does this come from? It's absolutely gorgeous. I want to watch all of those animes.
>>101971815you sound smart... for a dog
>>101971628Very nice. Catbox?
>>101961480bitch
Yeah so done some testing and for me personally on a 10400k loading the clip via cpu is simply not worth it. The clip text encode load itself takes 35 seconds every time I change my prompt, as opposed to 1.59s on GPU.
>>101971903>The clip text encode load itself takes 35 secondsyou mean prompt processing or is actually loading the model from disk every time?
>>101971903>The clip text encode load itself takes 35 seconds every time I change my prompt, as opposed to 1.59s on GPU.that's why I have a 3060 as my second gpu, can hold the t5 just fine like that
>>101971929I could rip the 2060S out of my sons PC but that would be mean. Plus I can't fit it in my case. When I do my next upgrade I will get a bigger case and mobo so I can dual GPU it.
Any poorfags using paperspace here? Are there any forge notebooks that can run flux yet?
>>101971962just use vast.ai
>>101971717I find the amateur photo lora on civit is handy for removing the "yep, this is AI" look that you often get by default with flux>>101971862Warning: Severe Joy Caption boomer prompting https://files.catbox.moe/0yh19m.png
RX 5700 XT AMDlet here. I just bought a 3090. Can I use the AMD as a second GPU for something? Is it worth it? I have 56 GB of RAM.
>>101971964paperspace's monthly fee is preferable to me, gpus are pretty much always available
>>101971969You have mismatched RAM or something? That's such a random amount.Also yes, offload the clip to your 5700 like this guy >>101971929
>>101971980Used paperspace for a year, it was annoying with SDXL model sizes because of the tiny storage, can't imagine with Flux
why is comfyu so slow when using the NF4 version of flux? Forge smokes it.. but Comfy kills it in everyother models.. PDXL, SDXL and 1.5
>>101972002Why are you using NF4 instead of Q4 though?
I seriously hope you studied the blade
>>101971994I stored models themselves in the temporary storage (ram) and loras on storage, it worked out. download speeds are very fast so it wasn't an issue to redownload the model for every session
>>101972009coz I have a 3070, only 8G VRAMq4 needs like 12 something right
>>101971982Can I mix and match AMD and NVIDIA like that?>you have mismatched RAMYes, I have added modules over the years and I have 64 + 8 or something like that.
>>101972042Don't mismatch your RAM. You will get better performance from 2 sticks that are the same as opposed to 4 sticks that are different, even if the total RAM is more.
>>101972017no, they're the same size
>>101972060I know. I needed amount over speed.
>>101972078DDR4 is pretty cheap. You should buy a kit and have the best of both worlds.
>>101970903okay so i did a completely fresh install, used the update.bat and moved all flux files onto the new install and i still got the same problem
>>101971967Alright thanks anon. Yes, I also use amateur photo lora myself for portraits, didn't realize it was also this good for luxury houses, architecture etc...
>>101972087I'll keep that in mind. But I just bought a fucking 3090.
>>101972071I didnt know that! Is it better than NF4 quality wise?
>>101972089try fp8 file or use comfy, seems the Q4 file is bugged or somethingmake sure it's Q4_0 not the K files
>>101972094yes it ishttps://reddit.com/r/StableDiffusion/comments/1eso216/comparison_all_quants_we_have_so_far/
>too lazy to make art the hard way so I proompt>too lazy to proompt so I copy paste joy caption proomptsI have peaked
>>101972093what are you, poor?
>>101972103grab from here: https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main
>>101972108Ill give this a run! thank you kindly
>>101972113I'm out of liquidity at the moment is all.
>>101972091I didn't actually test it til now but here is what >>101972111 would look like without the lora, it really helps.
>>101972135sell the house
Your delivery of bread has arrived...>>101972156>>101972156>>101972156
now this is what technology was meant for. to better mankind through innovation.
>>101971967>find the amateur photo lora on civit is handy for removing the "yep, this is AI" lookDoesn't always work so well depending on the gen
>>101972297>>101972319>>101972332these are great gens, what model did you use?
>>101972349midjourney for the 1344x896 init image, then made big and nice and errors corrected using Flux in 2 upscale stages using Ultimate SD Upscale nodesI'm finding it ideal because although the Flux's own txt2img artwork looks slopped, if you give it something non-slopped in img2img it respects the style extremely wellAnd because it's a large, smart model you don't need to tell it what it's looking at, I leave the prompt empty and it knows from looking at the MJ image what everything is supposed to beIt's a perfect combo because midjourney's own upscaler is shit, and their V6 model, while soulful, makes a lot of mistakes which Flux, being less soulful but smarter, can correct
>>101972465i'm going to murder you
>>101972488it's too late, you said you liked them and you can't take it back nowI'll die with a smug look on my face as you kill me
>>101969649fuck sake, i am literally falling in love with this chick i gen'd using the flux model. she's perfect.whats wrong with me i have a girlfriend already
Do flux quants work with lora yet?
>>101973805yes, they all work with lora now
>>101973447can we get a catbox please anon?
>>101974437"1girl, standing"
