Discussion of free and open source text-to-image models Previous /ldg/ bread : >>102000715 >Beginner UI EasyDiffusion: https://easydiffusion.github.io Fooocus: https://github.com/lllyasviel/fooocus Metastable: https://metastable.studio >Advanced UI Automatic1111: https://github.com/automatic1111/stable-diffusion-webui ComfyUI: https://github.com/comfyanonymous/ComfyUI Forge: https://github.com/lllyasviel/stable-diffusion-webui-forge InvokeAI: https://github.com/invoke-ai/InvokeAI SD.Next: https://github.com/vladmandic/automatic SwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed out https://rentry.org/sdvae >Model Ranking https://imgsys.org/rankings >Models, LoRAs & training https://civitai.com https://huggingface.co https://aitracker.art https://github.com/Nerogar/OneTrainer https://github.com/derrian-distro/LoRA_Easy_Training_Scripts >Flux https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnell https://comfyanonymous.github.io/ComfyUI_examples/flux >Pixart Sigma & Hunyuan DIT https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma https://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiT https://huggingface.co/comfyanonymous/hunyuan_dit_comfyui Nodes: https://github.com/city96/ComfyUI_ExtraModels >Index of guides and other tools https://rentry.org/sdg-link https://rentry.org/rentrysd >GPU performance https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html https://docs.getgrist.com/3mjouqRSdkBY/sdperformance >Try online without registration txt2img: https://www.mage.space img2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest sd3: 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
>>102003576Thanks for the bread baker
>>102003621>haven't used pony at all since Flux came outI can never go back
>102003588I would be embarrassed to make such garbage level gens after putting that many hours into the hobby
>>102003663Flux really did set the bar high for any competitor
>>102003677Post cut
PSA: LoRas with Network DIMM 16 and Alpha 16 can supports multiple costumes in Flux Dev via token activation and booru tags. >>102003405>>102003569
>>102003695I must admit, I like first one better. oh man lol>>102003663you'll lust for "tentacle sex" sooner or later.
Did anyone the T5 GGUF? If so, how does the quants compare to f16 (are they comparable or shit like the fp8 version)
How are the T5 GGUF quants? are they comparable to the fp16 version or shit like fp8.
>>102003873The results are pretty goodNot much of a difference so far
>>102003873>>102003881use Q8 or Q6_K, it's free:https://www.reddit.com/r/StableDiffusion/comments/1ewpwtp/flux1_t5_v11xxl_gguf_clip_encode_compare_gguf_vs/
>Is that true anon?>You won't disappear once I take my meds?
LoRA training. 1024x1024 or 512x512 (cant be both, it fucks things up because of buckets)Rank 16 or rank 32... or more?So many little levers to pull.
>>102003930Tuning hyper parameters is my kind of pseudo-science.
>>102003952You don't write good luck prayers in the script them comment them out?
After extensively testing the instant x flux union model, I can safely conclude that it is complete garbage. It will never adhere to the input and the outputs by simply having it on are guaranteed to be fucked up.
You guys are fucking weird man
>>102003725>tentacle sexI think a well trained LoRA can satisfy that. Besides, single GPU (24gb) finetuning is now fully available.https://github.com/kohya-ss/sd-scripts/tree/sd3Time to caption your dataset.
>>102003696*Smell
Fuck yes thank God for math nodes, I fixed my upscaling issue. Setting a +32 pixel buffer on the tiles helps a lot.
>>102004050even me?
>>102004050it's like a freakshow, it's the only reason to come here
>>10200393016 is pretty good. I crammed 5 costumes into 16 ranks. 768x768... Cause even my 4090 on W11 seems to be leaking memory to RAM on 1024x1024 and slowing things to a crawl. Maybe there are a few optimization options available to make 1024x1024 viable, haven't check all of them yet.
>>102004104ESPECIALLY you
>>102004124c'mon man..., don't be like that
>>102004124Me! Me! I wanna be weird!
>>102003621
>>102004132>ellipses followed by a commaThe weirdest of all.
>>102004156Something is wrong with your cat. He may have a syndrome.
>>102004158..., ?
>>102004156Does it mean I win?
>>102004156He just saw lots of weird things lately, don't mind the cat
>>102004201Cool outfit. Prompt?
dude get off my The Sims 3(tm) lawnwait why am i an insect creature? how kafkaesque of you...
>>102004123Using Ai toolkit or Kohya?
>>101997631 >>101996982 >>101998799 >>101999002these gens look awesome
>outputs randomly coming out as fried noiseis this a comfy update or something? was working fine before
>>102004293thanks anon, if you want to reproduce that you can use this lora and this kind of prompt:https://civitai.com/models/667307>y2k style cover art with a stylish girl drawing, y2k style text at the bottom: 'FRESH'
>>102004212i dont remember i made it with an a2 lora a while ago. i think i used a gradient
>>102004342
>>102004361
>>102004374
>Train LoRA at 512x512>Sample images come out with the quality of a 512x512 imageI don't know why I expected any different.
guys...https://civitai.com/models/668646/full-female-anatomy-fineart-nudity-flux-lora-nsfw-breastspussyassand-more
https://github.com/city96/ComfyUI-GGUF/issues/54#issuecomment-2300129278if your images went blurry with ddim_uniform scheduler it's probably because of this
>>102004445it's funny how easy it was to uncensor flux, like this took not even 2 weeks wheras for SDXL it took fucking months to achieve something remotely close to this, and this is just with loras
I want a workflow that generates a sdxl/pony image with a lora, then uses flux to generate backgrounds based on that image (via controlnet depth or something like that)would be neat, the only thing pony can't do is detailed backgrounds like flux. sdxl/pony still have all the best character loras (for now). I mean anime style, there are plenty of good flux loras for real people.
>>102004405I've trained a few now at 512x512. I'm getting good results and likeness but I'm noticing at lot of fine detail is lost. I deliberately chose images where her cheek mole is very prominent for my dataset and it's barely shows up
>>102004291Kohya SS, with GUI
Is there an efficient loader for Flux?
>>102004306they finally added sovl into flux.. the future is bright.. i close ma eyes..
>>102004474it's not going to fit together, it'll look green screen
Feature/Update News:[--fast] A '--fast' option has been added that can boost performance for 'float8_e4m3fn' on 'RTX 4000' or 'ADA' GPUs. (A recent version of torch is required.)
Since lora training seems to work so well, we can be sure that finetuning will work equally well or better, right? Like it isn't technically possible for a model to respond well to lora training yet somehow be untunable?
>>102004488Right, I'm looking into it now, it seems the GUI isn't caught up to the SD3 branch yet... or is it?It's documentation isn't as straight forward as AI toolkit, and all the discussions of the subject are full of the cerfucking guy.
>>102004509or another option, just take an image and use something like segment anything to make a perfect mask in flux.
>>102004512I cant wait for anime finetunes so Miku can play with her friends without a bunch of loras.>Miku Hatsune fighting Saber from Fate Stay Night with a celery stick
>>102004512lora and finetuning are functionally the same thing one is just on a mass scale
>>102004467and that lora actually works, its abit samefacey (but the creator wrote so) but he is already working on v2 ... and with a few tricks I actually achieved tentacle porn with just that lora
>>102004525>I cant wait for anime finetunes so Miku can play with her friends without a bunch of loras.this, exactly this, what flux needs is just more knowledge of characters/celebrities and we won't be far from perfection
>>102004519>Inserts himself into discussion>Spams his ugly fucking face>Subscribe to my patreon to see my findings!Hate this fag
>>102004512>Like it isn't technically possible for a model to respond well to lora training yet somehow be untunableProbably not. They're very similar in principle. I think the main thing is that it might be difficult to actually "improve" the model. The dataset sees pretty good, so whatever you add into in a finetune probably also needs to be immaculately captioned and curated as well. I think you can do little fine tunes here and there and it probably won't do much, but big fine tunes might be a bigger undertaking than before.
>>102004536
>>102004486Did you tag the cheek mole?
>>102004445>all front facing portraits>all the same stupid landing stripit's all so tiresome. training complex concepts isn't this easy.
>>102004543the worst part is that I looked up at his patreon and this fucker is making fucking thousands per month, that's insane, I live in a world surrounded by retards...
brainlet here, the clip encodes your prompts and then sends the conditioning to the sampler, right. Could you encode a bunch of prompts you like separately, save them, then later send them directly to the sampler without ever having to load clip?
>>102004556>>102004543I hate him so much it's unreal. He literally does nothing. He just hangs around developments like a bad smell. He contributes nothing, teaches nothing that isn't explicitly outlined in the documents, he just IS.
>>102004544And the inability to sell the resulting finetune removes commercial incentives, meaning that only rich autists will be able to afford training the vast number of steps required to substantially change the modelWhich means the first serious finetune that totally changes the model will probably be made by furries
Tell me again why I need a LoRA?
>>102004567What like load a bunch of clip converted prompts from the text encoder into a document or something and the fire them into the sampler at a later date?I don't see why that wouldn't be impossible, but I don't know enough about what's going on under the hood.
>>102004445>no puffy nipple
>>102004577People are screeching way to hard about the license. Until BFL actually comes knocking for their payment I call bullshit on them ever enforcing it.
>>102004590>What like load a bunch of clip converted prompts from the text encoder into a document or something and the fire them into the sampler at a later date?yes, just as you described it
>>102004604who's gonna take the risk? do you think people are willing to pay thousands for a finetune and then having the stress of possibly getting a lawsuit because they decided to get that money back through donation and shit?
>>102004614That sounds possible, in fact, I bet you could put the script for the t5 outputs from the text ecoder in the chat GPT, ask it to modify the node to send it to a text document in the form of a list and then make another node that takes that list as its input and you could even cycle between your cached prompts with a slider or something.
>>102004549I included "she has a cheek mole" in my captions for the closeup photos. Maybe I'll try WD14 tags next
>>102004630I think you're overestimating the cost of the fine tune and the potential risk to the uploader. There are other concerns around BFL getting their money as well including international borders and bad KYC on the part of the uploader. Assuming the uploader was intentionally hiding their identity, which I don't think is even necessary.
>>102004540thats the only thing it lacks, I want an anime finetune and a real/celeb finetune. loras are fine but what is neat about pony models like autismmix is you can get GREAT results even with basic booru tags. it knows what saber/mordred/nero are, even without a lora.
>>102004586a simple lora enables full lewds if you desire it
>>102004630Just lie about what the base model is... I doubt they could tell after a fine tune.
>>102004586I wish I was as easy to please as you.
>>102004656yeah, you want ton include a new character to your gens, with a greatly finetuned model that can be instant, loras are just a patch, you have to searsh for civitai, pray that your beloved character exists, then download it, then load the lora, then wait for the model to unload/reload to include the lora (that can take minutes), and often the lora is too strong and alter the whole flux's capabilities, and you can't stick too much loras at the same time it's getting slow and flux starts to shit itself, it's tne antithesis of fun, nothing will beat a model that already has the knowledge inside of it
it's neat how much the model seems to know conceptually, even a "world of warcraft ingame screenshot with UI":
>>102004630How does CivitAI do it? People can pay buzz for early access to Flux Dev loras. Buzz can be exchanged for USD. Both CivitAI and the lora maker are profiting from Flux Dev.
>>102004667how are you gonna do that? the finetune will keep the model's size, everyone will know it's still a 12b model, and only flux is a 12b model
>>102004667There's no other 16ch model like it on the market, anon
>>102004687Civitai is actually paypigging BFL for a commercial license thoughThe discussion is about amateur trainers who won't want to do that
>>102004687>How does CivitAI do it? People can pay buzz for early access to Flux Dev loras. Buzz can be exchanged for USD. Both CivitAI and the lora maker are profiting from Flux Dev.maybe a % of that money is given to the BFL labs?
>>102004679yeah, flux makes amazing mikus for example because it has knowledge of miku in the data: now if a finetune did this for all popular anime characters you wouldn't even need a lora for most gens. But loras are still very useful, it's like plugging in a character preset.
>>102004693If you use it commercially, I assume you're talking about setting up an API endpoint instead of just giving the weights.
>>102004688Those 2 red dots are gonna cost you three days my dude.
>>102004718if it's just making an API that is problematic, then the pony-fag would've worked on flux dev already, it's probably about the donations aswell
>>102004655>I think you're overestimating the cost of the fine tune and the potential risk to the uploader.we don't live in the SDXL era anymore, flux is a 12b model, it's gonna be really expensive to make a real finetune of it, the pony-v6 finetune cost tens of thousands of dollars, and it was just for a 3.5b model, now imagine for 12b
>>102004725kek, its about time we replace the mods with AI
what is the diminishing return point for steps in flux? 20? is there any benefit to a 50 step gen, default is euler but it's not euler a
>>102004758I think most people use 25-30 for euler. That's what I've seen from the workflows posted here at least
now we're gaming. not even a WoW aesthetic lora, just base model.
>>102004758>what is the diminishing return point for steps in flux? 20?I think it's 30, that's when you get consistency in qualityhttps://reddit.com/r/StableDiffusion/comments/1er3wt7/if_you_want_a_good_compromise_between_quality_and/>default is euler but it's not euler abecause the ancestral thing doesn't work on flux unfortunately
>>102004741We've yet to really see how flux responds to finetuning, but it's actually quite resource efficient so the price might not be as out of this world as you think if everything goes well.
>>102004634I only ask because it might make the process of using larger models on systems with less memory more efficient
>>102004771I hope so anon, and I agree it's true that flux needs less work to be good, SDXL base needed a lot of work, what flux needs the most is more concept and some NFSW training and that's pretty much it
>>102004783My hope is that you can train what you want into the model, throw the model out there and someone else will merge it back into their model until an ultra muttified model is produced.
>>102004794that was the process that created good models for SD15 at the beginning.. omy good old berrymix
rip and tear!
>>102004783the core of flux is very good so the gigabrain people working with it will make really good stuff with it. SD3 had fundamental flaws that they can't fix.
Goodness no! What have I created!
>>102004829I wanted something cute like this! Not that furry cancer!
>>102004848>yuria man of culture I see
>>102004848Here, I failed gen for you
Two half-assed kissing prompts on two different models
how do i actually do a proper hiresfix/2nd pass on flux? i turned denoise to 1 to test it and it just generates a blurry image rather than the oddly scaled images like 1.5. when i lower the denoise, the same thing happens: i just get a slightly more blurry version of the input
>>102004950I don't understand why people are still using Stable Diffusion workarounds on Flux...
>>102004950Flux works at any res out of the box because of its architecture. Although in practice, results may vary.
>>102004950I've never had issues with inpaint on comfy with flux. Forge was fucky.
>>102004950Don't use latent upscaleUpscale the decoded image with nearest exact/bicubic (or an ESRGAN model if you want to get fancy), then vae encode and run img2img at low denoise
How much benefit would I get from moving this from my sata to nvme ssd?
>>102005021I imagine model loading speeds would go way up but nothing else
>>102005021the model would load much faster for the first genafter that no benefit, unless you're swapping models a lot
>>102004950either this >>102004990or use SDUltimateUpscale .. I scale 1280x720 gens with SDUltimateUpscale using flux itselfpic related
>>102005046>Those fused arms.
Lora's on Flux are so much more effective than they were on stable diffusion. Flux is just so perfect in many way. Just makes me more pissed that Stability AI was trying to pass off SD3 as the next big thing.
>>102005048>Those fused arms.Love at its finest
>>102004800>berrymixjesus, that was a long time since.NAI v3, Elysium, URPM, ATSB....
>>102005074Bstaber
>>102005082Isn't that the pizza model?
>>102005089people said that, but I only ever used it for landscapesit was really good for them for some reason
>>102005089that was the mythical cstaber
When are we getting actually good controlnets, not the cheap shit that instant x is putting out?
trinart... My sweet
>>102005157why does this make me so uncomfortable to try to read/look at it's like uncanny valley or I'm having a stroke
>>102004543Jesus fuck, even if I filter him from GitHub just knowing he exists irritates me. It's worse than Nochekaiser
>>102004961>>102004974a flux image at 2k doesnt magically produce a bigger version of the 1k with the same seed. it's stupid to waste time generating at 2k+ base res only for the seed to be shit.
>>102005217>a flux image at 2k doesnt magically produce a bigger version of the 1k with the same seed.it doesn't, but the quality is still great at 2k, so why not making your gens there
>>102005217If you want to upscale then use a proper upscaler, not the hiresfix shit
>>102004446They do
Man, a fashion lora with designer clothes and more face variety would be sick just to make outfits and cool poses. Flux really enables way more creativity...
>>102004567Yea, this is possible. But seeing how t5 is quantized now the time it saves is negligible
>>102004990DAT that shit.
Can I set the outputs folder for saved images on comfy to a different drive? Every time I try to set one it errors. I don't want them in the outputs folder
>>102005298No I didn't recommend that for a reasonIf he's impatient enough to be trying to use latent upscale to avoid waiting for the extra vae steps, there's no way he's going to tolerate the speed of DAT
https://civitai.com/models/668799/game-assets-cartoon-style-3d-isometric-background-assets-for-small-games-flux?modelVersionId=748705that one is a win
>>102005321that absolutely make sense. I'm easing into it. the heavy ones are nasty tho.>>102005351very nice!
It just occurred to me how fucked Stability AI actually is. Basically anyone of any worth left, all they have is a handful of finetuners. All the researchers are gone. Wonder what they’re even doing at this point.
https://huggingface.co/XLabs-AI/flux-ip-adapterholy fuck that's a huge deal no? with this you just need a picture of something to make it work
Please gen literally anything other than Miku, I'm sick of seeing her obnoxious face, I'm begging you. Anything else. Every thread is spammed with Miku. I'm so tired..
>>102005413there's no much native characters than Miku that Flux knows...
>>102005412but can I run it with less than 16gb VRAM yet
>>102005445it's also somehow both undercooked and fried at the same time. just be patient, a proper ipadapter is going to take months.
>>102005445>ditch the original comfyUi repo?are you the same anon who didn't understand he had to pip install the requirements.txt from yesterday... be honest...
>>102005427Given you can run flux with 8GB VRAM using the gguf quant checkpoints and that these models are 2GB combined, I'd think yes.
>>102005412>Go to ComfyUI/custom_nodes/x-flux-comfyui/ and run python setup.pyI don't think it's that simple, it would be installing the packages on the default python shit, and not on ComfyUi's own venv
>>102005483>ggufwait, is this compatible with gguf quants though?
>>102005472No, he's not, that's me
>>1020035763060 isn't enough anymore
>>102005412>Download our IPAdapter from huggingface, and put it to ComfyUI/models/xlabs/ipadapters/*.I don't see a "xlabs" folder in there, should I create it?
>>102005626yes.
>>102005412having an ipadapter is a huge deal for fixing style if it works, since you can just show the model an image in the style you're looking forbut xlabs' controlnets fucking suck, so I don't have high hopes for their ipadapter being good
>>102005661>having an ipadapter is a huge deal for fixing style if it worksand for characters aswell, no need loras with this shit
flux does inpainting well too, here's a random inpaint of a random KCD image with a square blob:
>>102005529
>>102005683Thought you were bullshitting at first, that little HUD at the bottom looks so damn clean
>>102005676it's not as accurate, it sees only one perspective of the character
>>102005412wtf it's OOMing the clip encoder part
>>102005712I only did an inpaint on the mid-right area in a box. normally it's a forest area. but if you do WoW UI, it kinda knows how it looks.
>>102005714yeah, InstantID is better for example, I was able to use just one picture of anyone and it could work on every angle
6GB vram mafia bros...
Is latest comfy update still a bit broken for some?
>>102005412It's a shame quality suffers from the race to be first, then nobody else bothers to try. Like cool, it's ipadapter for flux, except it looks worse than 1.5 and take minutes per generation.
>>102004644Does your lora work with different poses and camera distances besides the generic portraits?
>>102005725anon... it still is making shit up, it only has one perspective to work witha lora can get all around detailInstantID can't know if your character has a huge scar on the back
>>102005744>InstantID can't know if your character has a huge scar on the backthat's too specific to be a redflag, 99% of characters have normal bodies
https://github.com/XLabs-AI/x-flux-comfyui?tab=readme-ov-fileipadapter readme, dling files now
>>102005786and even if it isn't perfect it's new, but controlnet functionality in general is one of the best things about 1.5/XL. When canny started working for SDXL models you could literally do one click edits, source image -> prompt/lora character
>>102005412Ehhh??? It's not even accurate to the original picture
>>102005661>but xlabs' controlnets fucking suck, so I don't have high hopes for their ipadapter being goodthis, the quality is atrocious
>>102005814
>>102005825xlabs always sucked though, so the "beta" thing is just copium
>>102005723Can you catbox that?
>>102005814you got a blonde anime girl with no girl in your prompt, it even has the details like a bow, that's still impressive
>>102005829it may ta
>>102005841>you got a blonde anime girl with no girl in your prompt,that's because I used the exact same prompt as their workflow, I guess that's how it's supposed to work?
>>102005412ya it works, somwhat. This was made with a picture I made in SDXL. It changed the style considerably. For comparison this is the original ... (next post)
>>102005839all I did was inpaint a box on the right part of the image with the prompt "restaurant", here is the source image: all default settings, denoise was 0.75 and default inpaint settings.
>>102005852anon I think you're retardedlike no offensejust the way you respond to things is retardedlike you don't quite get what is being said
>>102005412>>102005814>If you get bad results, try to set true_gs=2ok that's better, maybe using a character with no background was a bad idea, I'll try something else
>>102005857.. this is the original. Maybe it works better for photorealism.
>>102005860>all I did was inpaint a box on the right part of the image with the prompt "restaurant"Oh shit that is impressive then. Thanks for the heads up, will try
>>102005863>a retard calling others retardssuch many cases
>we are patching diffusion model, be patientis it patching the base checkpoint? is it fine for non ipadapter gens?
Error occurred when executing XlabsSampler:_scaled_mm_out_cuda is not compiled for this platform.do I need a new cuda or torch or something?
>>102005882My IQ is 124
>>102005871>Maybe it works better for photorealism.ok in the trash it goes kek
Yes these numbers are coming together nicely indeed.
>>102005922>124>He can't even count>D>>102005930oesn't realize 3 comes after 2 What a fucking unrepentant tard..
>>102005412>x labsFrom the moment Flux was released, they have be sprinting to be the first to release the most undercooked and shitty controlnets possible. It's almost offensive how bad their work has been.
>>102005945Wtf how did that quote of >>102005930get in there, I'm committing sudoku out of embarrassment
Everything is changing so fast.I very much doubt the way i had flux set up over the weekend is the most efficient method any more.Is there a handy compiled chart of what models work at what level of quality and speed chart floating around, did we come to a consensus?I use Linux on a 4060ti 16GB a 13th gen 13500, 64GB ddr5 and a fast m.2Speed is the priority but happy to give up some for maintaining the highest quality that fits the specs with wiggle room for Loras etc. No worries if not, just wondered if I'd missed some vital resources.
>>102005683>KCDbased
>>102005930kek, yea this seems to a work in progress at best:I tried to make her hair black - failI tried to make her ride a bike - failI tried to make her realistic - failI tried to make her be painted in picasso style - failI tried to make her a V sign - fail.. two more weeks.
>>102005968I'm not convinced the quant models actually help with speed at all, it's just as slow for me using them as not. I figured they were just to prevent oom.
_scaled_mm_out_cuda is not compiled for this platform.I dont get why im getting this error, I have cuda 11.8, torch is installed, torchvision, etc, files are all there. fp8 checkpoint, but error when I try to use the ipadapter.
>>102006004>I figured they were just to prevent oom.it's a huge deal for the vramlets, they aren't forced to offload some of the model into the ram, and that makes shit way faster
>>102005968the concensus is, shit's whack. gonna be a while before things calm down.>>102005892kinda soothing, I like
>>102005373indeed
Error occurred when executing ApplyFluxIPAdapter:'Linear' object has no attribute 'out_features'fun
>>102006004there's nothing inherently faster about them, they just compress the model to be smaller which helps a lot if you can't fit it all in VRAM and is useless if you can
>>102006027did you update comfy and that xlabs shit?
>>102006035will try, used to doing an occasional git pull but there are rapid changes I guessstill thisError occurred when executing ApplyFluxIPAdapter:'Linear' object has no attribute 'out_features'it dies at apply flux ipadapter node.
>>102006055im gonna do the clone repo stuff and do it fresh, might be something that messed up
>>102006069Error occurred when executing XlabsSampler:_scaled_mm_out_cuda is not compiled for this platform.do I need a certain CUDA version or torch, everything else for flux works
stop wasting your time with xlabs trash, nothing they make is worth usingthey are rushing out garbage because they want to be first and they don't care if it's actually usable at all
>>102006113this, 100% this
Does nvlink give you more available VRAM?
>>102005892>>102005915calming
>>102006142no, sadly nvlink doesnt work with desktop cards and torch, best you can do is use your second card to load the text encoder for some minder speed improve
>>102005727>gtx1060 wait forever*ftfy
>>102005860
>>102005727>not cpu forevercasual
>>102006223How long does a CPU gen take?
>>102006222awesome, inpainting blew my mind when I first used it and it still blows my mind how it can alter reality, basically. To do that without AI would take ages in photoshop.
>>102006231forever
>>102006068s-sovl
>>102006236yup. my cock gonna be so long.
>>102006231I'm not actually doing it but lol, it was 3 minutes per iteration when I tried. Maybe that could be improved down to 1 minute with the fastest possible RAM and a really good CPU?
>>102006259>Flux>cock
>>102006068which lora/style is this? I know it's animal crossing
>>102006285nm found ithttps://civitai.com/models/668799/game-assets-cartoon-style-3d-isometric-background-assets-for-small-games-fluxlooks like it sorta
>>102006259>>102006236flux is an inpainting beast. hands, feet, guns, you name it. of course, one thing it can't do, the naughty stuff. I inpainted an old (real) stupid image I had lying around with a cup and a pricetag and changed one letter, it even created the jpg artifacts. its scary good.
>>102006316it can do lewds with the topless loras I think, for regular gens they work, and if not you can just do that in sdxl/pony inpainting if you want
>>102006303cool, that'd be an amazing skylight desu
https://github.com/aredden/flux-fp8-apiLooks like he improved the fp8 speed, I wish it would be the same for Q8
sorry nogen because working on smut>>102006331yeah "fluxtastic" seems ok, havent had enough time to try the other ones yet. I ended up doing some inpainting with a sd15 model for non mutated nipples and bulge removal, works too I guess.>>102006303slick
>>102006303this would be a good gen if it weren't about some soulless psychiatric ward-white abomination that only a sheikh with shit tastes would consider tasteful
For the attention of horny anons across the globe, Flux can create nudes if you prompt it in the other languages it's trained on which is French and German.
>>102006345I'd buy it
https://civitai.com/models/668799/game-assets-cartoon-style-3d-isometric-background-assets-for-small-games-fluxDoom Eternal ingame screenshot, the main character is Miku Hatsune, she is dressed in a green suit of armor like doomguy from the videogame DOOM and is holding a large shotgun. The background is the first level of the FPS game Doom Eternal with fire coming from the ground and skulls filling the floor. The image is in the style of the game Animal Crossing.random doom prompt + animal crossing style + lora, magic
gday anons
>>102006468>attachment missing
>>102006426any French or German training was incidental and happened only with T5, not Flux
>>102006426very nice. gonna try me some honry later, in german. danke>>102006468BRO
>>102005727can flux run on 1060
Am I taking crazy pills or are all the things xlabs putting out simply not working? I mean that in the sense they actively ignore the controlnet input while making all the output gibberish.I think ONE of their canny models stuck to an input once. Aside from that it's all been shit.
>>102006494explain what you mean to a low IQ zambo like me, I thought the text encoder is intrinsic to the model used? If Flux used T5 in the training, wouldn't that mean Flux has horny concepts trained for german/french tokens?
>>102006527Run is a strong word. It can limp or crawl on a 1060
>>102006531Being correct is not important, what matters is being first and shouting the loudest, such is the nature of the grift
>>102006462oh! Miku met a villager!
>>102006541I mean yes it'd be a humiliating experience but for some reason I thought nf4 didn't work on older gpus. renting cloudgpu would make more sense I guess
>>102006545catcatbox? i always appreciate a good art render
>>102006561but they are friends so it's fine.flux is going to be a godsend for concept artists/modelers/people that want graphics btw, that's a perfect isometric mario.
>>102006567>nf4Nobody is using NF4, just variations of gguf.
>>102006539The words in different languages will be linked to a certain degree but the diffusion model itself was trained with English captions and T5 itself was trained for English.Where did you read that T5 knows French and German?
>>102006479Is this fine establishment hiring?
>>102006596The source of all known knowledge that's infallible, Reddit. A Redditor said so, which makes it the truth.
>>102006608redditors are retarded, I know because I am one
>>102006608What the fuck is German for boobies? Boobenßnell or something?
>>102006629gros seins
>>102006629Boobenbounsendammerung
>>102006426it can do that for English as well
Hey guys...
>>102006650turkman made it to 4chin, no shilling the patreon here sirs
>>102006539catbox or not real
>>102006629I asked it nicely for a little "aux seins nus, cuisses épaisses" and received thighs and tiddies
>>102006568https://files.catbox.moe/puyvjd.pngit's messy and you'll get a lot of errors about missing nodes probably
>>102006650I hate this fucker. It's always the same test images, he never tries something complex that would show his shitty loras are trained shittily.
>>102006578the power of friendship!it's kinda wild how you are only limited by creativity and prompts now, in the past you could try to paint something wild, but it would still take a lot of effort to make a finished product, this takes mere seconds to conceptualize. Technology is awesome.
>>102006664>aux seins nus, cuisses épaisseit may work in french.. just tried in in German.. no go .. got ugly german old women with bikinis
>>102006666thanks, i can figure it out
>>102006596"certain degree" is the key expression here. using french or german (those horndogs!) words just gives the latent vector created by the text encoder a little nudge in whatever direction. the model obviously can generate ass&tiddies.>>102006629brüste. titten. (rude) euter. (even worse, think cow), etc>>102006578this looks fantastic lol
>>102006668Is there a feature on github where the owner of a repo can block people from posting in the discussions. I shouldn't have to see a link to this guy's patreon when I'm reading up on the latest Kohya pull.
>>102006661https://files.catbox.moe/wiuwuk.pngIsreal
>>102006669>it's kinda wild how you are only limited by creativity and prompts now, in the past you could try to paint something wild, but it would still take a lot of effort to make a finished product, this takes mere seconds to conceptualize. Technology is awesome.the only limitation now is the lora, we don't have a lora on everything, I hope a serious finetune will add more concept to flux so that we can get everything we want without downloading any loras
>>102006677I forgot to include the starter image since it's a Flux img2img upscale of a small midjourney gen, picrel
>>102006697Flux nipples are kind of boring. What I liked about pony was the sheer variety of nipples. Flux is just little circles, and the LoRAs coming out have not helped that of all. Nobody here has heard of an areola.
I get topless output and thick thighs in english as wellI dunno what you guys are on about
>>102006697bwahahahaah those are not tittes nor puss
>>102006697tumor nipples
>>102006703damn maybe flux can generate that natively one day. midjourney remains winning for art, as always.
>>102006700I think the base model knows what animal crossing is but this lora was trained on graphics like AC and other games so it gives really good resultshttps://civitai.com/models/668799/game-assets-cartoon-style-3d-isometric-background-assets-for-small-games-fluxfinetunes with game characters or anime characters will be amazing, for pony there is a model named autismmix and it can do TONS of characters with basic booru tags. loras are great but you dont even need them a lot of the time.
>>102006697so it's the same big chewable nipples that Flux creates with English wordsyou fucking retard
>>102006697>https://files.catbox.moe/wiuwuk.pngyaaa okay I see what you meant but thats still flux nudity, I can do that to in good prompting in english, the problem is the nipples are borked and her pussy is just a black triangle, but you are right atleast with the french it goes easier to nude than in English (just tried with google translate of my prompt), you can do it in English to tho, but you always get thos pluff nipples and no genitalia
Miku Hatsune is in the game Legend of Zelda, she is dressed in a dress like Princess Zelda. The background is the first level of the game Legend of Zelda. She is talking to Super Mario from the game Super Mario Bros and giving him a high five and smiling. The image is in the style of the game Animal Crossing.
>>102006725yeah but midjourney's gens are small and their upscaler is shitso I'm using MJ to gen starter images and then making them huge and nice with that Flux workflowworks nice as you saw with the cat
can anyone tell me about align your steps sampler?
ignore derpy mario, miku as zelda is pretty cute:
>>102006767>miku as zeldaI'm like 99% sure that it's just Zelda and the Miku prompt got overpowered.
>>102006747I'm impressed all it takes is "The image is in the style of the game Animal Crossing."
Well GEE YOU GUYS I'M SORRY if everyone here is genning nudes with english tokens, for me they never worked and instead I got feet
Early morning bread, get it while it's fresh...>>102006777>>102006777>>102006777
>>102006273I'm getting 5 minutes/it for a 1 mpix gen. Must be busted to hell.
Miku Hatsune is in the game Animal Crossing. She is talking to Tom Nook from the game Animal Crossing. In the background is a store with a sign that says "Tom Nook's cocaine shop". The image is in the style of the game Animal Crossing.lol>>102006775im using the lora trained on 3d/isometric graphics too, it makes perfect villager types.
>>102006788you have to use something like "artistic nude photography" or "she is unclothed and nude also she is naked" .. like double tripple.. but if you wanna see what is possible with just one lora and flux check this:https://litter.catbox.moe/0jjboj.png
