Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101694073>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Kolorshttps://gokaygokay-kolors.hf.spaceNodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper>AuraFlowhttps://fal.ai/models/fal-ai/aura-flowhttps://huggingface.co/fal/AuraFlows>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>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
>early bake
>>101701184>146 images
can someone autistic enough give us a breakdown on all the new gen diffusion models, pros and cons
>>101701184is over 300 text msgs silly, learn your 4chan
>>101701058Dear baker, what do you use to make those OP collages?
>>101701235>literally any image editorwtf anon
took me too long to write these so>>101701015if you're having trouble with it not following style-related parts of the prompt, try dialing down the guidance to 1.0-1.5. the default 4 works better with short/low-effort prompts; lower will listen better if you're actually putting in effort.>>101701026i have covid rn so 0 but if i didn't it would still be 0 as i am violently unfit>>101701125>>101701152you both know what i'm going to say (can't talk about data etc)>>101701169We don't have anything to announce w.r.t. future models/releases at this time, but we're all big fans of open source here so it would be reasonable to assume that we'll be releasing more open source models in the future.However if you ask me *personally* i'd say it is extremely unlikely that we would ever release -dev under apache2.0. Regrettably we do have a business to run & allowing hundreds of copycats to spin up an API service that undercuts ours (since they didn't have to fund development the model in the first place) just doesn't make sense.again, personally, I do hope we'll be able to set up a simple flat-rate commercial license for -dev in the near future, but that's very much not my division and we have nothing to announce there at this time.
can flux do inpaint and how do i accomplish this in comfy (first time user because flux)?
>>101701235
>>101701237Sounds like lots of manual workI thought there might be a script or toolJesus
>>101701239>try dialing down the guidanceHow do I do that in this default flow, there's no guidance scale
>>101701239Reposting from the previous thread:>Can you please tell the folks at Replicate to add the ability to manually change resolution when generating images? They don't let us do more than 1MP with any of the FLUX.1 models right now, and we can't control step count for dev.This is really bugging me out, why does FAL allow this but not Replicate?!! Are they intentionally making your model worse to use? I want to generate Flux images >1MP with Replicate, but I can't.
>>101701239>i have covid rn so 0 but if i didn't it would still be 0 as i am violently unfitwaah get back into bed! thank you for all the answers so far .. one last silly oneCan you trigger safety_tolerance in dev somehow?
>>101701254you can do it with imagemagick but the command lines to do so are pretty arcane>>101701260uhhh good question. i think you need the advanced guider node instead of basicguider? i dont have it running in comfy atm now>>101701266I dunno, I'm not involved with any of the Replicate/FAL stuff and implementation details are up to them; I can pass on the feedback though.At a guess i'd say it's probably because they have a fixed per-image price & 2MP gens are significantly more compute intensive than 1MP>>101701275not a clue, I haven't looked at what it even does on our end but I do know we've partnered with a third party service for part of our safety filtering
>>101701239Has your team considered a TerDiT version of your model?
>>101701304>At a guess i'd say it's probably because they have a fixed per-image price & 2MP gens are significantly more compute intensive than 1MPBut how come FAL can make it work but Replicate can't :( Anyway, thanks, I hope they fix it. I can't use their discord because it requires phone verification.
>>101701260Add Fluxguidance node.
>>101701239When you say more open source models do you mean just image/video models or potentially music models as well? Would be nice to have an open alternative to Udio.
>>101701326Anon, music industry is the one that you don't want to get in with an open source model...
>>101701304No question. I just really like the model.
>>101701260either this >>101701316or use CLIPTextEncodeFlux, also got a guidance scaler
>>101701316doesn't exist but cfgguider does and that does the trick
>>101701307personally I think all the ternary shit is a fun toy but ultimately a meme, at least until someone actually goes and builds a native-ternary processor. it doesn't achieve much of anything that you can't do with just fp4/fp6/fp8 and has no real meaningful performance improvementDiTs are *kind of* like LLMs but they end up compute bound, not memory bandwidth bound, so shrinking the model weights (the main benefit of ternary) is much less of a priority than it is with autoregressive models>>101701326once again I must break out the "we don't have anything to announce w.r.t. future models/plans at this time" card, sorry anon
>>101701260actually do >>101701345 or your you loose the ability to even use CLIP (not that that useful in many cases, but you might want to) .. with your current workflow you can only use txxl5
>>101701356Update your comfyui, Fluxguidance only got added late yesterday.
flux can generate pregnant teens really well, top tier model
>>101701210pros: >really nice comprehension>great text rendering>less censorship compared to previously censored local releasescons:>awful stylization, likely trained on synthetic slop>incredibly beedy and hard to finetune
>>101701342thanks! happy people are enjoying it. whole team worked very hard on it for a long time & it's nice to see that work wasn't for nothing :)>>101701376again, lower your guidance to 1-1.5 if it's not listening to style prompts as much as you'd like it to
>>101701336That's the issue, that's exactly why we need an open source model. Just as with AI art, every song that comes out of a music model is a derivative (unless tuned, which wouldn't be the responsibility of model devs). Once the cat is out of the bag it can't be contained.
>>101701361>>101701345I will be honest with you, I don't know the difference between clip and txxl5
>>101701388neggles why don't you create an issue or reply in github to verify yourself?
>>101701376it lacks kino unfortunately >de3 gen
>>101701413clip: old SD15, SDXL style text encoder. It understands your prompt in very simple terms and works best in tags.txxl5: complex llm like text encoder, which is used by SD3 and FLUX which works very similar to Midjourney level of understanding prompts where you can write a more direct complete discription, with spatial awareness etc. you see on your own workflow you have a dual text encoder, you load both clip and txxl5, the way you routed it inot a TextEncod Prompt will just grab one of them, but there is a new custom CLIP Text Encode FLUX node that you can just place where your curent Text Encode (Prompt) sits, you got two text prompt options then, one on top for CLIP that you can prompt like always, and one below were you can write lovely prosa discribing very complex scenes and get insane results
>>101701462T. Sucks at prompting.
>>101701474.. and that one also has a guidance slider
>>101701437do i need to? i could tweet something i guess, tried to make a gist but the antispam says no when i post the link *sigh*
>>101701413pic related is what you want to use for FLUX
>low quality JPG CCTV photo of a small red circle drawn over an area of the sky with a little Pope Francis inside the circle
>>101701510can you just share a catbox with embedded workflow
>>101701388cranking the cfg down completely erases the character and i just wind up with a generic paiinting.
Found that pumping the guidance up helps the style, finally can make world of warcraft screenshots.
>>101701496just leak the flux version before it got sanitized, then we can verify. but seriously, good job guys!
>>101701523Damn making thumbnails is so easy now kek
>>101701474So txxl is for boomer prompts, got it.
>>101701483>nogen
>>101701530up????? neggles said to put it down......
>>101701537i wonder how much bandwidth i could waste by posting a magnet link for "flux pro leak" that's just a bunch of random-initialized tensorspeople would notice pretty quickly, sure, but how many people would download it before then?>a customer service counter at a store, behind the counter sits a seal wearing the store's uniform and a hat, a sign hanging from the ceiling above the counter says "WARRANTY SEAL"
>>101701525sure >https://files.catbox.moe/hravji.png>>101701545not just boomer prompting, also its what understands the text, has the spatial awareness and cand understand multiple subjects in a prompt all that that CLIP lacks... its the default mode for flux, with a simple prompt node you will only reach txxl5, with the flux node you can reach both text encoders
>>101701510Thanks I already found it>>101701525https://files.catbox.moe/sb95jj.png
>>101701483Remember Dalle is a very opinioated model. It has a really good aesthetic default style (which I deciphered while using HunyuanDiT), because it was RLHF'd to do so. It's basically a manga /comic hybrid type of painting. You can achieve those results easily in Flux, or at least the aesthetic if your goal is to replicate Dalle's style, but you have to know what to ask for.
>>101701563I don't know... putting it up made the images WoW.
>>101701563sometimes i am wrong about things. schnell needs a higher number than dev, tooYMMV as usual, these models are only predictable in their unpredictability
>>101701579>you can reach both text encoderswhy would I want to when boomer prompting is just straight up better?
>>101701579side note: you should really feed the same prompt to both text encoders (which is what the default setup does)also thanks to T5 it's now capitalization-sensitive!
>>101701604there were some observations yesterday that if you just enforce a style in clip in extra to the txxl boomer prompt it will have some significant effect way more than when you mention somehwhere in the txxl prompt "painted by Van Gogh" or something .. also its nice to have options to use both .. who knows how effective it might be to combine em, I am still experimenting on it
>>101701548I am AFK hence nogen, your prompt is literally just colossal and then play around with fantasy creatures, year, robotic, blueprint, anime or manga style, etc...
>>101701587>Remember Dalle is a very opinioated model. It has a really good aesthetic default style (which I deciphered while using HunyuanDiT), because it was RLHF'd to do soOnly the vivid style (default for Bing Creator/Designer), Natural DALL-E style doesn't have that and you can make it output absolutely anything, especially if you have Azure endpoints with disabled NSFW checking/most prompt filtering. It does require better prompting though.Also, DALL-E has same issues with nipples as Flux (tested on those endpoints). Also see https://catbox.moe/c/lfnwjt, the prompt was "DeviantArt" + jb to force the LLM to not rewrite prompts.
>>101701594Do negatives work properly on dev/schnell? Will you add them to the APIs?
>>101701620>>101701635Alright well I just put them bot to inputs and hooked up a primative because I can't be fucked. Turning down the cfg helped a lot though
>>101701644I'm sure at its core, Dalle is the best model in existence. But outside of that you get dogged for simply asking for a completely SFW picture of a girl doing anything.
>>101701741>doggedSee, you're using bing creator, but my experience with dalle is 100% API only, where you can actually get natural style, + no-NSFW checker endpoints from Azure.
>>101701575How come companies never provide controlnets with their models. Would make the so much more attractive from the getgo.
>>101701753goy must pay for ze api
so what are the reqs for running flux?schnell can fit on a single 24gb and for dev i need one with more than that or can i somehow utilize both of my gpus?
>>101701760I understand why you had to take the tripcode off for this response.
GYATT
>>101701620>side note: you should really feed the same prompt to both text encoders (which is what the default setup does)Based on my experience this is not always the best approach. Pic related, left is same prompt in both, middle is empty clip, right with "pixel art, retro graphics" in clip. The text suffers, but the style is much closer to what I wanted.
>>101701594Did you see https://desuarchive.org/g/thread/101685374/#101688297Is it real?
>>101701770if you go back a few threads there were some anons that used a second GPU to load the encoders, while using the main GPU for the model in a 3090+3060 combo .. that said for running it without swaping you will need 24GB .. with swaping I seen some ppl have it running on as low as 12GB with forever wait times, but it will work
>>101701673I think they do? AFAIK we have support for them in the API but I'll admit I've not looked>>101701741dall3 is just a really big SD1.5/SDXL-style model with an unhinged VAE and GPT4 prompt preprocessing, it's thoroughly mid>>101701753*shrug* you'd have to ask research, my guess would be we figure the community will have that covered soon enough>>101701770if you can run schnell you can run dev, they're the same size, schnell is just a 4-step so it's faster>>101701788hm interesting, that does make some amount of sense desu, there's a reason we kept clipL in there and didn't go full T5-only
>>101701753Hunyuan literally did though?And IPAdapter on their roadmap as well, soon to be complete. They recognize what the communty needs. Hopefully Flux can keep up.
>>101701304>imagemagick collagesYeah that's terrible usability indeedI'm baking sometimes on a different board, so I'll probably make a little scriptI'm not gonna arrange pictures by hand every time like some caveman >>101701237 >>101701244
>>101701376Stylization is lacking even with low guidance (1.5).Lower right is Flux, the adherence is fantastic. Lower left is img2img with the same prompt and a SDXL Checkpoint.Bottom left is what I would expect from a model that understands style tags to generate. Top are with Loras and IPAdapter applied. Everything you add and details change. But it's the best solution for stylization.I'm trying different guidance settings, but my impressions are that it can look like a photoshop filter. Maybe there is a sweet spot or better prompting might help but I don't see it giving a similar style without extra help (lora or finetune).
>>101701795kinda used to long waiting times, had to run a shitty 2060 before upgrading so i know the painand is schnell that much worse than dev? I'd rather try some mad gpu science than have to regenerate a model 20 times
>>101701811>>101701826 (me)ah i see, my badi guess ill give both a gothanks anons
someone post a screenshot of the new official bigma news
>>101701790i had not seen it, and wish i still hadn'twould be pretty trivial to fake so without the original supposed output image i'm inclined to call bullshit
>>101701861So you never saw anything close to that output during testing?
>>101701644Is this something you can pay for, or something you scraped that's supposed to be le-secret? I wouldn't mind renting it for a week or so to do comparisons. Can you try something like "Zelda in the style of Picasso?" only Midjourney seems to get it right so far>>101701824Yes it seems tough to achieve a hard-specific style that doesn't fall into a metastyle.
Flux can't be fine tuned
Has anyone successfully gotten negative prompts working with Flux yet?
>>101701744And it's completely worthless because it's SaaSI am sick of "safety" when is someone gonna have the balls to just drop a fully unsanitized model.
>>101701862SAI shill cries out in pain as he strikes you
>>101701811>dall3 is just a really big SD1.5/SDXL-style model with an unhinged VAE and GPT4 prompt preprocessing, it's thoroughly midnow we're getting to the crux of itbig model = kinounhinged = kino
>>101701882>Is this something you can pay for, or something you scraped that's supposed to be le-secret? I wouldn't mind renting it for a week or so to do comparisons. Can you try something like "Zelda in the style of Picasso?" only Midjourney seems to get it right so farI can give it to you, sure, I scrapped it (but you can disable those filters on Azure yourself if you're a company with a valid usecase) but there is still some prompt filtering sadly, and some prompts get rewritten even if you force JB. The NSFW checker is completely off though.I'll see if I can get such a gen, but yeah, DALLE prompt filtering is really spergy about copyrighted characters.
Flux is a meme, it can give you incredible images but without things like loras and finetunes it's essentially as useful as Dalle3. They just released this version to bait investors and attention, but won't release any way to train it (the open models are distillations and you need a custom scheduler that has not been released)
>>101701874nothing like that in any of the testing i was involved with, no>>101701884this is false, lora training works at the very least, finetuning dev should work fine, schnell will require some new tricks but nothing revolutionaryi think part of the problem is that i guess nobody's worked out the loss function for SD3/etc properly yet?
>>101701898I am confused about it to, on release night comfyUI-anon was rambling here that it doesnt have negative prompt support cause cfg is always 1.0 .. then other anons claim it has, no also >>101701811>I think they do? AFAIK we have support for them in the API but I'll admit I've not lookedsaid it does.. but atleast on the replicate site it doesnt .. so where are we at?Does FLUX support negative prompts or not?
>>101701923Odd, because I usually don't have a problem with a lot of copyright characters especially really generic ones like Mario/Zelda. I just want to see what the outputs are for a couple of artstyles like Picasso, Yoshitaka Amano, HR Giger, etc
>>101701884>>101700861
Creator of SimpleTuner creators of OneTrainer have stated that creating fine tunes for flux is not feasible. CEO and founder of invoke says it's impossible.Astralite says they will never work on the flux dev, because of license restrictions.And flux schnell is few step ugly distilled model.Flux is not going to happen.
>>101701950Yeah, OpenAI API DALLE is different from Bing one is this regard, and Azure dalle is also a bit different from OpenAI API DALL-E and Bing :) It's a clusterfuck really.
>>101701590looks really cool, prompt/catbox?
>>101701981the creator of simpletuner is a batshit mental case that spends all day shitty on every local project that isn't xer's garbage SD 2.0 finetune
Now that I look at it again I think I've seen that dead baby before
>>101701989in game world of warcraft screenshotThat's the prompt
>>101701981I knew it, too good to be true. Fuck those people
>>101701950https://files.catbox.moe/0w02x4.jpghttps://files.catbox.moe/6oi04l.jpghttps://files.catbox.moe/ftnodt.jpghttps://files.catbox.moe/m75z91.jpghttps://files.catbox.moe/80oy4l.jpghttps://files.catbox.moe/lryyzl.jpgMind you, I didn't really spend much time on the prompt, and it's Natural style after all, and the API rewrote like half of the tries. I'll play with it a bit more.It really sucks for DALL-E to be like this, it's a fucking amazing model.
>>101701950Trying Amano in Flux completely fucking slops it up. I have a feeling this was trained on an amalgamation of Midjourney outputs.
>>101701981These people don't get to decide what happens and what doesn't.
>>101702018It's sort of there but looks more like some digital crystal glass filter or something, lacks that painted look.
So what was SAI spending 8 mil a month on?
>>101702033Nigger stop getting false hopes, you need an scheduler that isn't there. Even the CEO of flux said sohttps://github.com/black-forest-labs/flux/issues/9
>>101702046yeah the slop LLM generated "digital illustration", will change that
>>101701946supporting negative prompts isn't really a model architecture thing, it's an inference code implementation thingi know for a fact that it is *possible* to use a negative prompt with flux but we've not exposed it in our API because it's just not necessary, haven't found any benefit to having it when you can just add more detail about what you *do* want in the normal prompt>>101701981skill issue
>>101701813How is the hunyuan model anyway, slop or kino?
>>101702060>haven't found any benefit to having it when you can just add more detail about what you *do* want in the normal promptit's just easier in a lot of cases to add to neg instead of writing even bigger positive
>comfy's fault that negs won't work Many such cases
>>101700861>one of the people who made flux*>*i mostly keep the GPUs going brr but I moonlight as a researcher and sometimes have good ideas>>101701239>>Why can't it draw nipples>>obviously the question is whether you're training on cunny>you both know what i'm going to say (can't talk about data etc)>>101701099>Can't answer any questions around dataset or training resolutions because I literally don't knowhow would you have zero clue about the dataset and training res if you "mostly keep the GPUs go brrr" and sometimes contribute to research?the resolutions used is basic info when selecting images for a base model, or finetune. what about adjusting batch sizes according to the res to not oom, binning them and training to different degrees to images of different res/style/quality, et c?it can hardly be a black box for you, if you actually managed the training process
negative prompts are absolutely needed. you cant just tell it "dont generate a hippo" in the positive prompt...
>>101702050>the public Flux release seems more about their commercial model personalisation services than actually providing a fine-tuneable model to the communityThis. There is no fine tunes coming, because it's not feasible in technological sense and even then, the cost would be crazy. It's not a community friendly project.
hii, does anyone know about MMD to AI gens? like these https://youtube.com/shorts/s52NZLjuZm4?si=aOi46rNAuZJDHkmG
>>101701981>>101701981>>101702050Why would I listen to him/them/faggotself?
here's some vivid dalle, yeah it sloppifies the style sadly, but it's just such an easy switch :((https://files.catbox.moe/m774c8.jpghttps://files.catbox.moe/pgy3h2.jpghttps://files.catbox.moe/0ks637.jpghttps://files.catbox.moe/rnvacu.jpghttps://files.catbox.moe/t2gbpc.jpghttps://files.catbox.moe/oll6z4.jpghttps://files.catbox.moe/98l9ib.jpg
>>101702050
>start A1111>do SDXL gen with controlnet, takes like 3s>next gen with same parameters prompt no changes takes 3 minutes
>>101702075It's because it's bullshit, baka.
>>101702060Sometimes there isn't a good positive to add, due to biases in the modelCaucasian white people are often not labelledIf an Asian shows up in my gen and I don't want that, I just have to use negativesAnd yeah I know that sounds racist
>>101702060>it's an inference code implementation thingokay thanks, now we can bug comfy with implementing something Also when will FLUX.wurst be released?
>>101702126VRAM leak
>>101702091>the cost would be crazyNo it's notI've LoRA'd 70B LLMs, the cost is not the problem, this is just a kid journey that uses your hardware, absolutely no customization. This is yet another ideogram-tier service with the marketing strategy of releasing some unusable scraps to the public.Just ignore it
>>101702129>And yeah I know that sounds racistNigger, did you come here straight from reddit?It's a fucking preference.
Can someone explain me how SDXL Lightning on Replicate is $0.004 per image yet Schnell is $0.003 despite being a much bigger model? Is the compute being subsidized?
>>101702072true when we were stuck in 75-token CLIP town (using torch.cat() bullshit to get past 75) but is much less true now that we have big boi text encoder>>101702075I maintain the infrastructure and platform, not the training code. I don't spec training runs or model architectures or any of that, I just troubleshoot crashes and network fabric and performance issues, diagnose and replace faulty nodes, maintain storage infra, etc. Hell, I still don't actually know what the whole model arch looks like because I haven't needed to know>>101702074>>101702130well not exactly. our reference code doesn't implement it and he's working with what we gave him>Also when will FLUX.wurst be released?what, like a flux cascade model? for why?>>101702126>A1111found your problem
>>101702061Prompt following def. better than SDXL, but not as good as Dalle or Flux. Styles are kino, knows anime and manga really well. Also knows characters. There are finetunes in the making E.G. https://www.shakker.ai/modelinfo/87e2cc2169934523a2ff82fb12e7206b?from=feed. Check out some genshttps://imgur.com/a/hunyuandit-0vrZEn0Also just look in the archive here for Hunyuan filename to see some of my other gens.
>>101702162>spend investor capital on initially selling a product cheaper than you'd like to, to attract customersnovel discovery, anon
>>101702060Easy example is using a anime character tag such as a touhou and wanting them in a different style. Try it out and see how much of a pain it is to wrangle the inherent style bias.I'm saying this here in hopes people will listen
>>101702148I was trying to preempt the obvious pozzed replyShould have written>inb4 racistinstead so as to not offend le 4chins insensibility
>>101702104People who contribute in this space have some weight behind their text. How would you know better than them? You have never contributed anything of value. You simply consoom and coom to AI images.Flux will not have the ecosystem around it that people hope. If you like what you have now, then that is fine for you.
>>101702181it's two blocks*
>>101702166>ad hominemsngmi
>>101702060>because it's just not necessaryNot even for controlling colors, hues, that sort of stuff? I like to put certain colors to neg and let it run only few first steps. No need to wrangle so much with photoshop/gimp
>>101702185read the last comment on your thread numbnuts and sit the fuck down
>>101702166>well not exactly. our reference code doesn't implement it and he's working with what we gave himow okay.. well lets see what automatic makes out of it, did you give him access to?>>101702166>what, like a flux cascade model? for why?no, sorry I was making a joke! your team being German and all .. wurst would just be a good model name. Tho a cascade model would be fun, but I guess totally out of scope of what you actually do.
>>101702166How is your company associated with comfyanon?
>>101702208my guess its just early access so comfy can implement it before release
>>101702166Is there a specific order in which we should prompt flux? Like first the character, then the background, etc?
>>101701851>bigmaanon has all but forgotten...
>>101702189
>>101701851>>101702230
>>101702215Isn't it already implemented? They've linked it on their Hugging Face page.I came to think of it because he mentioned a1111 above.
>>101702246local won
Started the simple webapp to try out flux dev/schnell with NSFW checker disabled again: https://jack-june-mlb-finishing.trycloudflare.com/pls dont share with other generals, some of them are schizo. You can gen up to 4 images at once. If any of you want to play around with fast schnell or dev gens, feel free.
>Just in: Black Forest Labs refuses to deny that FLUX was trained on cunny
>bigma will be good >trust me anon
>>101702162my guess is it's a combo of "sdxl-lightning pricing is older and hasn't been updated" and "Unets are a huge pain to run and use way more flops than they have any right to"fr fuck unets dawg all my homies hate unets>>101702208we have a partnership with comfy org which, amongst other things, lets us ensure that comfyui and swarmui have day-0 (or day-1 when mcmonkey manages to sleep through the entire release day) support for our models>>101702220not really, it's kinda position sensitive because CLIP only gets the first 75 tokens, and it does have somewhat more emphasis on earlier parts of the prompt than later ones for similar reasons, but mostly just be detailed and specific. claude makes a good prompt upsampler too>>101702205>did you give automatic1111 accesshe gets access to the code when we publish it>wurst jokeah lmfao valid, my b
>>101702282>claude makes a good prompt upsampler tooyeah 3.5 sonnet is a godsent LLM
>>101702274Every model is. There's no human looking at every single trillion images to look for cunny.The only way to filter it is to use an algorithm that is trained on cunny
Can you at least answer how much synthslop went into the dataset so I don't have to wait for the paper to confirm the obvious?
>>101702272niggas why when I share this the first thing you do is prompt cunny?
>>101701241bamp
restarted the tunnel by accident, also disabled the VPN checker that I had:https://singer-immigration-consist-chen.trycloudflare.com
>>101702282Thoughts on>>101702178
>>101702298>I don't know anything about the dataset>Can you at least tell us X about the dataset?
>>101702305I tested what it would produce with single word "loli" and then I tried "anime style loli". Because the thread said it was trained on cunny. And for loli it did produce little girl, which was funny.
>>101702282>he gets access to the code when we publish itloool.. a1111 btfo
>>101702179Racism is in now after 2020.
>>101702326it can easily generated naked children (when I was running the earlier version of this webapp today, I saw those, it sexualizes children easily), and schnell can generate necrophilia with children
>>101702246discord invite link?
>>101702333Yeah. Any model that has children and nudity in it can always combine those and create nude children.There is no way around that, unless you remove all children, or all nudity or nuke the model in some other way.
>>101702352Yeah, it's just that since Flux is so good, the generated images also look very good, not like base SD 1.5 heavily mutated stuff.
>>101702350https://discord.gg/rde6eaE5Ta
>>101702282why not try to unsure 0-day a1111 as well? more the merrier, no?
>>101702356Everyone here implies SD1.5 couldn't do it, too bad Lexica took down their SD 1.5 images, it was 100% possible just 1 in 100 gens not in 1 shot like Flux.
>>101702302You obviously refused to answer. So it wasn't misleading.
>>101702377These are comfy's old coworkers they will suck his dickJust the way it is
oh here we go again.the "uhm we should only allow drawings that the police and "society" says is ok" retard again. sime time like yesterday. >>101689009using some fucked up shit as bait/excuse even though thats not even the level he would ban.saw a flux guy in here and just couldnt help himself i guess. >aktschually its not a drawing its a artificially generated image!! you are minimizing problematic output! fuck off retard
overlaying fingers looking good desumust have had lots of mirror selfies in the dataset
>>101702406The fact you're seething proves the point, enjoy prison.
i guess i shouldn't expect anything faster than 90 sec per gen on schnell as a 8gb vramlet
>>101702385>a channer makes an SD model trained on cute and funny and makes a company out of itbased
>>101702313thanks it works schnell is so fast
>>101702416i am not tolerating your bullshit dude.ignoring control freaks like you is what got us where we are now in the first place.somebody posts some weird ass loli dumpster shit and the hero "anon" mysteriously comes to the rescue and the solution minutes later. same today. couldnt help yourself but link the deleted post again. you are fucked up in the head man. lol
More ingame screenshots, I love emI'm trying to find a prompt / guidance value to generate good Zelda 64 fake screenshots
>>101702449Impressive
>>101699236>>101701153I love your fucking aesthetic, that shit is so fucking good. I'm a hobbyist making a retro FPS out of my Doom mod so I was trying to nab some aesthetic inspiration.it's kind of tricky but since I stole your proompt I'll share what I beat out of itStylized 480p pixel-art retro screenshot of DOOM, Duke Nukem (1993-1994) rural urban scene, pixelated early low-tech 2.5D FPS textures, low-poly objects, sprites, first-person shooter perspective with visible weapon. Low contrast. Weapon: dark green-camo submachinegun. Foreground: asymetrically on the right: sandstone formations of varied shapes, shrubbery and trees. Centerpiece: a wide shallow valley in the ground with three natural staircases leading down to it, buildings by the side, the valley is a cemetery with many medieval gray tombstones, grimy branching cobble paths, tumbleweeds, many lush trees. Background: dilapidated brown brick buildings with shuttered dark-gray windows. Brutalist concrete monolithic oppressive buildings. Environment: overcast desert sky, dusty atmosphere. Air: sandstorm. Style: high-resolution detailed pixel art, baked in simplistic shading, procedural textures, jagged edges. Key elements: retro FPS charm.
>>101702317Last call before I go
>>101702448It's already been ruled that it's illegal. There is no country that won't arrest you.
>>101702461There are countries where real child porn is legal
>>101702385>prove a negative or you are guiltyFuck off, retard
>>101702126>>start A1111use forge or reforge
>>101702464Name them
>>101702461>muh police justice power fantasiesneck yourself, fascist
>>101702469
>>101702471I'm stating facts, feel free to be the one that sets legal precedent.
>>101701940This is a cope
>>101702465That's not proving a negative. He simply refrained to deny that they had trained on it.
>>101701530how high are we talking here?
>>101702493he's just samefagging, like with the bakes
>>101702493his name was quite arousing, was i just horny the whole time?
do you reckon blackforestlabs will work on video gen?
restarted the webapp, enabled pro, it might take a very long time to generate or even timeout
>>101702313uh... what makes it uncensored? are models on hf censored?
>>101702515didn't they say they plan on releasing a text to video model?
>>101702493seemed legit enough, atleast we know about negative prompts now
>>101702515https://blackforestlabs.ai/up-next/
We got outshitposted by reddit
>>101702493Doesn't matter if he was real or not because he provided nothing of substance.
>>101702461>It's already been ruled that it's illegal. There is no country that won't arrest you.There it is again. Again you are pointing to your precious authority to give you your moral compas.Its not actually, I'm in japan.You in bongland or something?Pic related would put you in prison there. I bet you support that as well.If I would ban stuff that is sick and I dont like to see I would ban violence and gore on top of lots of porn shit, I dont think its healthy for people. Its all or nothing. Anyway you are exposing yourself again already. People can see through this bullshit.
>>101702525oh, my bad. replicate has nsfw checker enabled on the playground (you can disable it in the API), but apparently the hf demos have it disabled so you can gen 3d cunny right on HF
>>101701239>We don't have anything to announce w.r.t. future models/releases at this time, but we're all big fans of open source here so it would be reasonable to assume that we'll be releasing more open source models in the future.>>However if you ask me *personally* i'd say it is extremely unlikely that we would ever release -dev under apache2.0. Regrettably we do have a business to run & allowing hundreds of copycats to spin up an API service that undercuts ours (since they didn't have to fund development the model in the first place) just doesn't make sense.>>again, personally, I do hope we'll be able to set up a simple flat-rate commercial license for -dev in the near future, but that's very much not my division and we have nothing to announce there at this time.Are guys willing to give licenses for people like Pony who make more Anime and NSFW aligned tunes, in exchange for a % of the money he makes, or would it look bad?
>>101702545i'm kinda dumbare these models filter-free?https://huggingface.co/Kijai/flux-fp8/tree/mainor any local model does not have a filter built-in?
>>101702562>or any local model does not have a filter built-in?this, but models could instead have their dataset pruned, which Flux did. So it doesn't know genitals, and nipples are horrible. And generally sex scenes/cum/etc won't work.
>>101702545You won't get porn out of it but you can generate a swastika
>>101702573you can get softcore porn out of it, nude women from behind with their ass for example
>>101702542>in this news story, an anon residing in Japan was arrested after finding terabytes of photorealistic sexually explicit images of children on his hard drive
>>101702584
>>101702594pedophilies sounds cute <3
>>101702578that's really not porn, it's artistic nudity, it's really incapable of anything titilating
/ldg/ slowly turning into /sdg/ with all this kiddie talk
>>101702546highly doubtable that any corp. that wants to make an international name will work with pony crew on a license, the stuff you can make with it and especially the realistic derivates is guaranteed bad press, pony crew will have to work with schnell, cause apache 2.0 .. but didnt they just announce they will work on auraflow next?
Can anyone make flux generate hairy armpits for anime girls? I'm just doing experiments, and so far it cannot.
>>101702607Don't worry they're pushing hard for their perma, I've never seen people seethe so hard about their right to gen certain images.
>>101702499Regardless, I hope you understand that it will always possible to make questionable content because AI generation is able to interpolate between concept, so you'll always be able to get to cunny by combining enough related concepts.Trying to prevent this as part of the model is not only futile, it is going to hurt everything else as well and will make it strictly worse and drive it towards being useless to everybody.This has happened so many times now and I'm tired of people jumping at shadows instead of retards being told to fuck off and use digression if they are moving towards obviously illegal tier.Ultimately it is whether you think you can explain to a judge that it totally isn't illegal and if you should be distributing it online to begin with.
>>101702610>pony crewI'm sick of these fuckers dictating what we can and cant use with their massive porn dataset.
>>101702607
>Give case for why negs are needed due to wrangling token bias with a clear example. >ignored Why are they like this?
>>101702594Care to post the prompt for this "anime girl says" template?
>>101702584think so to .. also the man has covid, let him rest, probably has fever nightmares of this later
>>101702639anime girl with speech bubble that says "text"
>>101702632Wait, are we back to pretending that negative prompts aren't worth it?
>>101702671See my above post I gave a clear example for why it's needed>>101702178
>>101702685I don't disagree, I still can't believe auto almost didn't implement it because people thought it was a meme
>>101702667Well that was quite literal
>>101702696It's foolish and actually frustrating we are still having this conversation
>>101702707anon flux understands natural language
>>101702639I tried to get her to say the prompt with no avail but it's basically >>101702667 with the addition of "cute smug"
>>101702671We just had the answer in the last >>101702060 thread. The inference code that black forest provided to comfy as well as their API does not include support for negative prompts, doesn't mean the model does not support it, but in the current implementation it does nothing.I agree with >>101702685 tho that negative prompts are valid and useful. I hope for an implementation on that, no idea how that is done tho. So it remains to be seen when an anon is fed up with no negatives and writes one.
>>101702741I don't know how knowledge for negative prompts has already been so quickly forgotten, it's not like SD shipped with negative prompt support. Negative prompts are a community feature.
extreme prompt comprehension
pass space as the prompt to explore the latent space
>>101702741Damn, next they'll say prompt editing is worthless because you can just add more words.Some years ago I watched a video of some art guy on Youtube that said AI generations are going to become less interesting once the tools start improving and I'm starting to think he was right(I unironically think one of the 1.5 merges are still much more interesting because they know obscure art styles and artists names)
>>101702778yea, there is some truth to it, I keep my old sd15 models, they are like artifacts that will create some "vintage" AI eventually that modern AI will not be able to do, still I think flux is step forward on understanding of prompts etc., but I love my SD RNG gamble sometimes when you hit something really cool just by the wonkiness of tag prompting, model mix and other factors that you could not control, that you now can control
Can someone explain guidance scale to me? When should I set it lower/higher? What about styles?
>>101702298i can forgive a little sloppa when its comprehension is SOTA
What's a very cheap way to check if images contain naked children (only realistic photographs)? Thanks beforehand.
mfw I see flux requirements
>>101702848
>>101702839wdv tags + florence and looking for keywords
justice for vramlets...
>>101702848Distilled 1-bitnet flux-dev <2GB when?
>>101702889I actually hate the hawk tuah meme. It's fucking gross.
>>101702839pipe it through google/amazon cloud storage or discord, that should work
>>101702050Not based at all
>>101702911latency would be too big
>>101702854Nice
>>101702313a pro option just appeared, but errors out
>>101702966yeah idk why it's failing, maybe replicate is having issues again, their pro deployment is highly unstable
>>101702900It's peak...
>>101702902
>>101702624Still refusing to answer the question, huh,
>>101702454
>>101703011are you using schnell for this? because dev gets text much better
Do weights do anything in flux?
You can make semi-nude photos on HF spaces of any subjectCan use more steps and tags to make nipples sharperkissing and groping also worksno other... restrictions
>>101703047Yes I was
>>101702875this gives adult women loli/child tags tad too easily
>>101703173it's the combination of tags you're looking forand you would want to err on the side of caution if you're providing some sort of service
>>101703198nta, i just wanna stop people genning realistic 3d cunny on my flux webapp
>>101703192
>>101703020It's doubtful they used illegal images but if you have topless boys in your dataset you already having everything you need for topless girls.
>>101703165>no other... restrictionsHoly shitThis is gonna get pulled from there so fast.
this is impossible for flux to do:A close-up photograph of a human hand against a plain background, palm facing the viewer. The fingers are positioned from left to right: thumb fully extended, index finger fully extended, middle finger fully bent inward touching the palm, ring finger fully extended, and pinky finger fully bent inward touching the palm.
>>101703226yes you can generated realistic cunny on HF right now, they're really happy about that I guess
>>101703233just use SD wth controlnet nigga
>>101703210You posted a link to 4chan I don't know what you expected. But honestly if you wanted to be safe you'd have an LLM judge prompts for safety and then you would wdv+florence and have an LLM judge the results for safety.
>>101703223pictures of adult women with small breasts, topless guys, anatomical sketches and photos, beaches, saunas, artistic nudity as paintings and photos, et c. it's all there.
>>101703210
>>101703247so flux is shit?
>>101703233No shit
>>101703233it's a dumb fucking request
>>101703253inability to make an image of >>101703233 vs "shit" is a strange dichotomy
Hmm
>>101703277I don't know why people are so confused about how these models work. They can only generate images based on the captions they're trained on. There is no caption that is written like that and there is no LLM translation layer that normalizes autistic prompts (although that might be a good idea).
>>101703115
>>101703313What? The model should be able to generalize and be able to represent novel concepts. Bending fingers isn't something new, and hands are not either, why is it just generating a normal hand? That's just retarded.
I'm new to this and just got comfyui to be installed.I'm trying to download Flux but I have no idea what to get.I know the base model is in transformer which should go in comfyui "checkpoints" folder, and vae in comfyui "vae" folder, but about the rest ?Do I need them?Is there an easy way to download everything without going link by link on HF?
>>101703370it's not a matter of generalization, anon. it's a matter of how much information from your text input that can be used for compact model conditioning.
>>101703370EARTH TO ANON AI IS NOT INTELLIGENT>The image captures a close-up view of an outstretched human hand against a neutral background. The fingers are splayed, with the thumb extended to the side, while the other fingers are slightly curved inward. The skin appears smooth with a natural pinkish hue, and the visible lines and creases suggest the hand is well-used and has experienced various activities. The lighting is soft, highlighting the contours and textures of the hand without casting harsh shadows.Be smarter
>>101703391But one of BFL devs literally said that their text encoder is smart and I should just prompt in natural language?
>>101703394holy shit it's like when you use your brain prompting...
>>101703384https://old.reddit.com/r/StableDiffusion/comments/1ei6fzg/flux_4_noobs_o_windows/https://old.reddit.com/r/StableDiffusion/comments/1ehv1mh/running_flow1_dev_on_12gb_vram_observation_on/
>>101703384Check out this guide: https://comfyanonymous.github.io/ComfyUI_examples/flux/
>>101703215prooompt?
>>101703419anon its literally the most basic prompts in existence, can you not describe things directly? a fucking gpt-4o mini can do this
>>101703401It's not smart you fucking retard. It can't generalize from phrases that aren't in any of the captions. You can't write prompts that aren't represented in the captioning. Be smarter.
>>101703401there's no negation there. there's still a cap on how much different details a model can take from a conditioning source, especially without frying.
>>101703436You're telling me that such an advanced model doesn't know something as simple as finger names?
>>101703429i guess i cant coz all my gens holding guns are cursed
Um bros... why did they include tay tay in the dataset, don't they remember what happened last year?
>>101703450are you using dev or schnell? schnell will ofc be worse
>>101703233>>101703253>>101703370>>101703401>>101703439that's bait
>>101703457SD 2.1 base
>>101703462It's not, I'm genuinely confused.
>>101703453eating != sex
>>101703439It associates text with color blobs. If the text is not in the dataset with the correct color blobs, it does not learn. It also learns outputs, it does not actually learn what anything means. It is not a human, it is not logical. It summons color blobs from a noise cloud based on text conditioning.
Has anyone experimented with flux upscaling? I found that euler karras with 0.85 denoise gives decent results at 6 steps, less decent results at 10 steps, and doesn't work at all at 5 or less steps. My rig is a fucking snail so I can't experiment much, does anyone have a better combination of sampler and scheduler?
>>101703488try this, https://x.com/AIWarper/status/1819470305963958754Let us know how u go
>Hyper-detailed screenshot of DOOM (1993-1994) ported to PS1 and N64, side view of urban scene, low-poly graphics with pixelated textures, CRT screen effect, first-person shooter perspective with visible weapon. Foreground: '80s sedan parked along sidewalk, rust-covered. Background: row of dilapidated buildings with shuttered windows. Environment: overcast grey sky, gloomy atmosphere. Style: high-resolution pixel art, early 3D aesthetics, Phong shading, procedural textures. Key elements: pixelated gore, retro FPS charm. Additional details: scanlines, color bleeding, dithering effects typical of '90s console graphics.
>>101703522
>>101703453
https://fluxpro.art/ pretty good site with constant flow of new flux-pro images
>>101703603all your prompts are instantly public and there's no way to delete them, beware
>>101703612
>>101703603>login with google accounthow about no
>>101703618>>101703650im not saying to use their service. its a nice resource for flux prompts
>>101703673>can type just about anything (that's not retarded) into it and it follow what you ask faithfully>need to copy prompts
>>101703682it doesn't understand styles, artists, etc well at all. especially for things like anime.
>>101703500He's using euler normal with 20 steps, I tried it with 6 steps and it looks about the same and takes the same time as euler karras.
>>101703693Enjoy it for what is.
>>101703682There's plenty of things that don't work and it still struggles with multiple subjectsSo prompt ideas are still useful, to know what even works
This morning I woke up estatic with flux.Now I'm lukewarm on it because it's legally difficult and physically impractical to train at all.Like, what's the point?
>>101703808Legally difficult if you wanna sell your stuff. If you do it for fun, its no problem at all. And the schnell variant is totally open, but there you are correct its difficult to train schnell.But I guess porn anons wanna make $$$ so they will go for another model to finetune
Why is flux so retarded with prompts like this?>a renaissance era traditional drawing depicting a woman with short purple hair in a pink ball gown dress. Painted on canvas with visible brushstrokes. AcrylicsIt doesn't get the style at all, even with lower guidance. fucking SD 3 medium gets it just fine
darn it doesn't know shibari
Why is flux so good?
>>101703870>typical photoslopwow so good
>>101703836but Shnell kinda sucks :/
>>101703847are you using the correct flux prompt node? "CLIPTextEncodeFlux"
>>101703879can you show what you get with my prompt with dev?
>>101703877guess the coomers hope will be AuraFlow >>101701120
>>101703897The retard could've been training a 3B Pixart model this entire time and right now have a good prototype model. As said before, Pony v2 is DOA.
can some anon try this prompt in dev, wanna compare to pro modelhttps://fluxpro dot art/prompts/clzdy8wv800aujlveygh6p638
>>101703922theres a link to a webapp abovel iterally
>>101703922
>>101703886>a renaissance era traditional drawing depicting a woman with short purple hair in a pink ball gown dress. Painted on canvas with visible brushstrokes. Acrylicsyes sure but the problem is your prompt, you want a painting but you ask for a drawing, that will confuse the model ofc
>>101703986a painting is a drawing though?
>>101703977
>>101703977thx anonNoticeably worse than pro it seems
>>101703416>>101703417Thanks.I'm installing it on ubuntu but what's the point of swarm if I already have comfy? Just another GUI?I understand that : - the base model is flux1-dev.sft -> unet (why not in checkpoints?)- the vae is ae.sft -> vae- the text encoder/clip is in the dev folder but the guide says it's in a whole other depo ( https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main). Why? Shouldn't they be in text_encoder or text_encoder2?A few other things:- where are all the arguments when starting comfyui? Any page available listing them?
>>101703998holy shit are you stupid?
>>101704005maybe better in 16:9
>>101703998but drawing has a different status in the model, it will reference all the comic and anime drawings, and not think of a renaissance painting, renaissance drawings are very rare (think Dürers hands) .. this is your prompt with>a renaissance era traditional painting depicting a woman with short purple hair in a pink ball gown dress. Painted on canvas with visible brushstrokes. Acrylicspainting instead of drawing
>>101704044Okay, I understand thanks. But where are acrylics and visible brushstrokes in this image?
>>101704044don't argue with him, no one says drawing = painting>produce (a picture or diagram) by making lines and marks, especially with a pen or pencil, on paper.
My prediction: in a week /sdg/ and /lmg/ will have less than 10% of posts in threads about Flux
>>101704053renaissance paintings are very detailed and fine, they actually tried to hide brushstrokes, so this is actually kinda wild for a Renaissance painting .. you maybe looking for impressionist instead. Also acrylic paint is an invention from after the first world war . .ca 1920-1930 .. you are looking for oil paint instead
>>101704066how did u get the CCTV footage look anon, prompt? All my gens have been slop
>>101704076Flux will have cultural relevance because it can do memes and actually production ready images.Have fun with your coomslop though. 2 weeks ago both generals were dead.
>>101704104Use dev, it just works with the most basic "cctv footage of" shit, schnell just doesn't get it
>>101704076what will replace it mr oracle?
>>101704118nothing, people's excitement will just die down when they will realize that there are no tools, no one is probably going to do a big finetune, etc. it'll be a fun toy like dalle, just with no restrictions, nothing more.
12 min for a 1mp gen ToT
>>101704128It'll never get old producing reaction images.Have fun in your coomer generals.
>>101704053>a impressionist era traditional painting depicting a woman with short purple hair in a pink ball gown dress. Painted on canvas with visible brushstrokes. oil paintingThis what you are looking for?
>>101704162I'm not into cooming to image gen at all, anon.
>>101704165more like this
>>101704166The only people seething are coomers
>>101704181idk, I just think that novelty wears too quickly nowadays.
>>101704185That's a personal problem
>>101704172lool.. thats modern art style, thats about 1920s painters stylepic related is how Renaissance art looks very accurate, nearly hyperrealistic, what your are looking for is like a sloppy Edward Hopper style, realism or pop-art
>>101704202you're talking to someone who thinks drawings are paintingshe has zero understanding of art and hasn't even bothered to do basic research
>>101704213>you're talking to someone who thinks drawings are paintingsStop talking shit about me, I realized my mistake, I'm ESL and not an artfag so I genuinely didn't know
>>101704213thats why I am trying to teach him for the lulz .. maybe he gets it, maybe not, does matter for my karmic score
>>101701388Assuming you're legit, thanks to the team there for this release. Really stirred up interest all around. If you're not authorized for public statements you might also want to tone back details on who you are and what you do to lower internal conflicts.
>>101704231anon if he was interested in learning he would have done basic researchbut I'm excited to watch you spoon feed him and get frustrated
>>101704221You know, when generating images, it's probably important to use Google first if you don't know how to use words correctly.
>>101704221dont worry.. I got that, you know now, art styles are complicated, ppl study that and make a living just looking at paintings and discussing in what when were how with what etc. they were made, thats not your jobalso here you go you can go from this prompt and improve:>a pop-art era traditional painting depicting a woman with short purple hair in a pink ball gown dress. Painted on canvas with visible brushstrokes. flashing colors, acryl
Here it is...>>101704253>>101704253>>101704253
>>101704273nta, im getting this from the replicate, is there some issue with replicate or something? guidance is set at 1.5
>>101704357I used 35 steps euler, guidance 1.8, maybe its abit luck to, pop-art is widely different, some are sloppy brushstrokes, some work are high detail, I would probably rework the prompt, but I wanted to get it as near to anons orginal one