Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101764165>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>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Kolorshttps://gokaygokay-kolors.hf.spaceNodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper>AuraFlowhttps://fal.ai/models/fal-ai/aura-flowhttps://huggingface.co/fal/AuraFlows>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
Blessed thread of frenship
>mfw
>>101770367Next challenge - realistic looking cum. Impossible.
>>101770422lol give me a few hours let's see what is possible. with schizo prompting, you can do anything
>>101770437Good luck. No amount of ambiguous words like "cream", "goo", "semi-transparent liquid" and so on helped me.
>>101769863Update on this.Nothing works anymore, im getting errors out of the ass ive never even seen before. Guess i need to reinstall it all again.I dont even blame AMD for this one, never update kings.
>>101770512sorry for the dogshit artifacts I merged sd3 and flux
>>101770512try:(snot), (pus), (messy eater), (white industrial goo)
just woke up, what did I miss?
>>101770721awesome
>>101770683>I merged sd3 and fluxabomination
>>101770683why does it have AI faceAaaaah!!!
>>101770721>what did I miss?We can crank up the cfg up to 8 without any issues now
>>101770738
>>101770606The HIP SDK should have that file, doing a cursory search. What happens if you search for that DLL in where you installed the SDK, do you find it? Also, for not finding DLL dependencies, you should be using Dependency Walker (https://www.dependencywalker.com/) to figure out if every DLL missing is accounted for.
>>101770820how?
>>101770820>speed plummets down into oblivion>"without issues"
>>101770838>how?DynamicThresholdingFullhttps://files.catbox.moe/haqdtd.png
>>101770422
>>101770845cfg had always the effect of speed decrease, are you new? every people who used negative prompt before used cfg and therefore had their speed halved, no one complained before
>>101770846whats all this disconnected stuff?>XYZ Plot>select Node inputs
>>101770935some testing shit nodes, you can remove them
>>101770860What words did you use? Is it consistent?
>>101770838I saw you post this pic yesterday I actually saved it, can you maybe gimme the prompt? it goes so hard
Drink Flux Colaeven if the bottles aren't always the same size
>>101770964Not really consistent, a lot of them look like candle wax. I have 2 4090s, offloaded the clip model to one of them and have the fp16 model on the other, so i bulk generate 12 images at this resolution at a time about every 1 minute or so. Maybe 10-15% of the images are passable.prompt is "A photo of a blonde girl, she has watery and runny white liquid jizz on her face."
>>101770947but damn cfg at 8 generates at a snails pace.shit took like 4 minutes for this image kek.>>101771034here brohttps://files.catbox.moe/anxy7y.png
>>101770833Ehh?????hiprtc0507.dll is not there
>>101771071thanks man
>>101771071>but damn cfg at 8 generates at a snails pace.you don't need to go that far, if you can get the same result with lower cfg go for it
>>101771077kek
>>101771109so where is the sweet spot?
>>101771152I have no idea, I'm actually making a XY plot between guidance and CFG to see where the magic starts
>>101771171thanks for offering your gpu cycles for science anon! looking forward to the results
>>101771201this, doing gods work
>>101771050cool
>>101771201>>101771207oh sheesh
初めまして、カーミット様。
>>101770512Mayo kinda works.
>>101770863we use CFG even without a negative prompt
>>101771255kek
>>101771257cfg when different from 1 halves the speed whether or not a negative prompt is added.
how she drivnig like thismaybe a cool racing game
>>101771078I looked up the instructions. It says the following.>IMPORTANT: If your GPU is higher/newer than a RX6800, then skip Step 3IF your GPU is below a RX 6800, you need to skip to Step 3.>Install AMD HIP SDK 6.1.2 from here: https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html>If your GPU is below a RX 6800, you need to do the following steps:>3.1 Install AMD HIP SDK 5.7.1 from here: https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.htmlI'm guessing 1.) Your GPU is lower than a RX 6800 so you used the wrong HIP SDK and 2.) When you upgraded your driver, it no longer works because the ROCm install doesn't recognize that driver but that is weird because ROCm really shouldn't be tied to the driver you use based on what I know in Linux so might be unlikely. Do 1 first and then only then do 2 and download the driver you had before.
>>101771296yes, that's what I saidyou, with your ESL, implied it was only used when there is a negative prompt
>>101771328>you, with your ESL, implied it was only used when there is a negative promptwhere did I imply that?
uhhhhat laest i figured out the prompts in filename thing
>>101771318conflicting tokens, whenever you stumble upon something like this, prompt around it. "from the side" or "head outside" and such
>>101771339>every people who used negative prompt before used cfg
>>101771354where's the lie? to use negative prompts you need to use cfg, yeah
>>101771212thank you
>>101771364I didn't say there was a lie...oh ESL anonie, you're too dummy and cute to be mad at*pats your head*
>>101771387?
>>101771400*pats your head even harder*
what do you think these guys are plotting?
>>101771428the protocols
>>101771442this guys like "FUCK ITS HAPPENING"
>>101771428Are these guys like catholic clerics?
>>101766903thx I don't have a good graphics cardI have an old HD5670
>>101771479diff anon but those gens are really cool, how long does it take for each one?
>>101771322I have a 7900 xtx and use windows.And idk man it looks like its trying to call for the older version of HIP ver 5.7? But when i have that installed, it doesnt work at all.I read the instructions correctly, because it worked before.
>>101771490infinity, he is not using a HD5670 to gen with any model, it just ain't possible
>>101771506the GPUs we would have if they won the war bros... 1488 gb vram...
This is my new flux gf guys, please say something nice about her.
>>101771511isn't it possible to gen cpu only though? or use a shit gpu with the cpu doing most of the work?
>>101771520
>>101771522>please say something nice about her.Hey babe, you look like belle delphine, can I drink your bathwater?
Ipad Pro M4 running FLUX schnell, 4 steps, 100 secs per genApple won, bigly
>>101771549>25s/it
>>101771543Never mind ......
>>101771550dangerously based
>>101771495The only other thing is that your ZLUDA build is outdated. From https://github.com/lshqqytiger/ZLUDA/releases/, I am seeing a v3.8.1 that is from 3 weeks ago and and v3.8 right below that explicitly removes support for anything below ROCm 6. That probably may be the only reason why ROCm 5.7 is being sought after.
>>101771511>it just ain't possibleSays who?
>>101771564Faster than your RTX
>>101771490like 30 seconds with gradio webui (link from here) or a bit longer on huggingface
>>101771455neat
What's the best FLUX model I can run on my 16GB card? My internet sucks so I don't want to download a huge model just for it not to fit
>>101771522
>>101771836There's model with merged vae or something like that, should be easy to find
>>101771428
>>101771603Okay, well this is new.At least its not exception errors.Reinstalled ZLUDA and HIP SDK.just for the record:AMD Driver ver.24.7.1Python ver.3.10.11HIP ver.6.1.2ZLUDA ver.3.8.1
>>101771906Based Karen
>>101771857>>101771918I like the style
>>101771882I'll look around, no worries. I'm just glad it's not completely off the charts. Seeing such huge model sizes sent shivers down my spine. I run a 6GB SDXL model no problem, but FLUX is like 16GB at least. Scary.
New to SDXL lora training, is there any consensus on what learning rates should be set to? I downloaded a few models off civitai to check their metadata, and they're all wildly different.
>>101772011read up on prodigy optimizer and try that one first, it just werks.recommending this niggas settings for the optimizer:https://civitai.com/articles/3105/essential-to-advanced-guide-to-training-a-lora
I merged sd3 and flux
Can flux generate legible text in foreign scripts? I haven't had any luck yet with Japanese or Cyrillic.
>>101770820Doesn't this make it slow as fuck?I just woke up too.
>>101772047Neither of them produce tits this big
>>101772057using cfg halves the speed, it was always like that with that parameter yeah
I miss Ran
>>101772062>two cigs
>>101771922It found your APU at device id 1, probably on your Ryzen CPU based on the fact it's some weird RDNA2 variant on desktop. 7900 XTX is gfx1100, try using device-id 0 instead.
>>101772063my man you have no idea who you are talking to
>>101771606Me.>>101771609bait
>>101771935
>>101771836both flux dev and flux schnell are the same size and have the same vram requirements, you can run flux with 16GB VRAM if you load it in fp8>>101772084two lungs
>>101772164Several people in these threads have been running it way less VRAM.. just takes a while
>>101772251my shit crashes on comfy every time i try it
>>101772072Is it always half unless you use cfg 1.0? i.e is it the same using cfg 3.0 as it is using cfg 8.0?
>>101772251Yeah because it resorts to swap when it runs out of VRAM, per the CUDA sysmem fallback policy by default. I haven't tested Flux personally, just downloaded and have it ready to start testing tonight since I was busy this entire week and last week until today. You can also reduce VRAM usage using FP8 for VAE and the text encoder and the unet with the options ComfyUI provides. I haven't tested but traditionally, you could also use token merging with the tomesd node in the for_testing ComfyUI custom nodes and put it right between the model loader and when it gets to the sampler and increase the ratio to use less VRAM too which worked not only for Stable Diffusion but some other models too.
>>101772294did you --lowvram
>>101772369ya speed just doubles regardless how high you set CFG
https://new.reddit.com/r/StableDiffusion/comments/1emi1j9/opensource_amd_gpu_implementation_of_cuda_zluda/>a based gentleman wanted to help AMD by making Cuda compatible with their cards>AMD sent a ban notice to himIf that's not a sign that AMD is a controlled oposition, then I don't know what else to say
>>101772397>doubles*halves
>>101772392would that even helpi don't even see anything get loaded into vram it crashes before that happens
>>101772421post workflow
>>101772447bruh literally the glass bottle workflow to load the schnell model https://comfyanonymous.github.io/ComfyUI_examples/flux/
>>101772406Why would you link to Reddit, you mouthbreather, especially when it's news and it's old news. Use the original link to the website that broke the news.https://www.phoronix.com/news/AMD-ZLUDA-CUDA-Taken-DownAnd actually get the fuck out of here, we don't need to see more of you vermin bottom-feeders here who can't contribute positively to the culture and posting atmosphere who just are going to repost shit there once something useful here is posted.
>>101772505zam
>>101772397Pretty shit, I can barely run it as it is on a 4080. Interesting experiments but I don't feel like it's worth it for general purposes.
>>101772474might be a dumb question also but do you have the latest comfy commit? also just try --lowvram just to see
>>101772537i updated todayill try in lowvram just for you honey
>>101772526thank you
>>101772172>>101772062
>>101772568have you been at this for 12 hours?
>>101772526for my use case it's interesting, with cfg = 1 there's no way I can change Miku's skin, I thought flux was just unable to do something like that, turned out that cfg unleashes its prompt understanding to the max
>>101772537insta crashed.even the terminal kills itself
>>101772505I doubt any open source dev would ever take legal action against open source code but i am almost 100% certain you cannot just go backses on open source license agreements?
>>101772576yes
>>101772598respectI just woke up
>>101772598Thank you, king
>>101772090I KNEEL
>>101772582that probably means you don't have enough RAM or pagefile/swap space. make your pagefile/swap bigger. I get by on a 4GB card with 16GB RAM with Schnell. Blistering 125s per iteration.
>>101772626Should try amfetamine + vodka combo for ultimate art
>>101772626>>101772649I can barely get any sleep since flux came out, I just wanna Proooompt.feels like the early days of SD but better.
>>101772672(loaded as fp8 on checkpoints already converted fp8, even if you have the fp8 e4m3fn launch flags, it still has to load the full fp16 checkpoints before they take up less RAM/VRAM.)
third flux lora dropped on civitai, this time chinese style slop:>https://civitai.com/models/629898/flux-lora-littletinies?modelVersionId=704235
>>101772584Read the wording carefully.>The code that was previously here has been taken down at AMD's request. The code was released with AMD's approval through an email. AMD's legal department now says it's not legally binding, hence the rollback.This wasn't a legal threat, he basically lost the legal guarantees from the fact that AMD's legal department had approved his release of the code to open source and then retracted it. He basically did a preemptive move to prevent Nvidia from rolling out their own legal action against him after he lost AMD's legal protection with the code, basically and there is no way he can publicize the code any longer. Anyone else can pick it up and fork it since it was released with the orginal license but risk having Nvidia ram a DMCA in their mailbox.But I am just going to level with you and everyone else regarding this topic. CUDA compatibility was always going to be a slippery slope because it's a proprietary standard from Nvidia and isn't open source and thus is subject to their whims and desires per how Nvidia licenses it even if it's for compatibility. While it's a great term solution short to medium term, it can't be the long term solution. We actually need to see more software that isn't CUDA being written even if Nvidia is far ahead of everyone else.AMD's ROCm isn't a solution either because it also isn't a standard and is only open-source. The only solution that I see working with a slim chance and that I have tethered myself to is Intel's strategy with SYCL, a Khronos standard and the successor to OpenCL, and extending that to the UXL Foundation with trying to get everyone on board with it. Unsurprisingly, AMD and Nvidia are missing. It is my opinion AMD is hurting themselves not trying to write a frontend tying this to ROCm and I am hoping at some point AMD comes around to it.
>>101772568
>>101772714cool, that's for flux Dev, I'm scared that they will all focus on Schnell because of the licence, I don't want that, Dev is the best model we have
>>101772714why that bird leaking bean juice?
>>101772731regular basic bitch lora makers won't care but anyone with real money won't use dev unless they're backed by donations
>>101772726
>>101772731It's ok I will train a coomer lora on an a6000ada when the tools are a little bit better.
>>101772746You didn't read the dev license.
>>101772667Glad to see you getting it to work, it seems. I haven't used ROCm in over a year but it seems like similar errors and etc. from when I was using it, albeit a lot less convoluted than the Linux errors I saw with Vega 64. AMD still has a long way to go, there is no reason why ROCm shouldn't work on Windows at the moment natively but it is what it is.
>>101772793based Coomer Chad
>>101772749
>>101772798dev license is non-commercial use and you're in a grey area with some guy training models for fun and getting monthly Patreon donations
>>101772814>>101770274here's where the spam posts go
>>101772829>yuria man of culture I see
>>101772090>>101772667>go through the effort of updating GPU to play BFV because of error>errors out SD instead but can play>spend all day fixing SD>BFV errors out with earlier error, but SD worksLessons learned;>never update>any product made by EA is actually garbage
>>101772867>playing EA goy slopdeserved
>>101771039poison drink>>101772052yes, it can make legible Japanese hiragana but I don't speak it so It's probably nonsense words. I haven't tried prompting something specific.
>>101772881Honestly and BF4 actually works yet BF1 and V dont so you know what it was deserved
>>101772882>Our competitor's product is poisonMade me laugh
>>101772867>never updateNever update thinking it is a quick thing for alpha/beta software which is what ROCm is despite AMD saying otherwise. And yeah, EA's programming is kinda shit but also AMD's drivers too on Windows. Hopefully updating SD made it faster too so the effort was worth it.
>>101772961>updating SD made it fasterIt did actually, thanks.
>>101771857prompt?
>>101773009it's just a new oil painting lora on civitponyXL and a character lora
>>101772793>coomer loraI assume the best we would get is "1girl, spread anus, hairy asshole", because for actually complex scenes the whole model would need to be re-trained
>>101772714yaaa.. well I tried it.. its shit, people obviously need to figure out how to make good loras for flux still, hardly can get it to make anything
I'm getting lots of booba without even asking for it, the weird half-censored nipples are really noticeable.
>>101772974
>>101772672>>101772700got it to workaaaaand its just a blank fucking image lol!
>>101773181Do you have image previews on? Was the image black the entire time or were there colors and shapes before it went black. With black images throughout the entire process, typically, I find a restart fixes things but I don't have a Nvidia GPU. If it is colors and shapes before it goes all black, then yeah, it's something in the generation process going wrong.
>>101773233got it to work
>>101773318grats .. now make us some gens
Inpainting in Flux is dramatically better.
>>101773112nofap is easy when you have a pussy
>>101773407I haven't bothered because inpainting is comfy is nightmarish
>>101773441SwarmUI makes it a little better. But honestly even without inpainting the results can be great
>>101773376will do
>>101773441that hasn't been true for like a year now. You can draw a mask within a load image node and pass the image + mask straight to a sampler.
does Flux have negative prompt?
>>101773508nta but the issue i have with it is that it doesn't consider what is behind the inpainted maskso if i do it at half denoise i get a grey blob.a1111 doesn't do this.
>>101773517when you use the DynmamicThresholding node you can use negative prompts if you raise CFG above the reccomended cfg == 1 and set mimimc cfg == 1, but the generation time doubles then
>>101773553You're doing something wrong then, I've never had that issue
So, is Flux open source in the same way normal SD is? Like, are there people currently working on different checkpoint for porn (similar to pony)?
>>101773606can you provide an example workflow then?
is there any way i can get decent flux gen speed on a volta card?
>>101773633they are on the fence about that, the horsepussy model maker (pony) is bitching about license again cause he can't make money of it cause while FLUX.schnell is completly free and open source, the superior model FLUX.dev is only free for non-commercial use and the master model FLUX.pro is closed source, ppl still will make loras for it (already are) but if and when we see a coomer finetune remains unknown, but never underestimate the thirst of coomers for new toys I guess
oh shit i got nipplez on flux i thought it wasnt possible nvm they turned into fucking grapes
>>101773640just load an image, open image in mask editor, and then send to a sampler. I usually use a mask ksampler which is an impact node, but I think these ones are all default nodes
>>101773824>set latent noise maskthe node that has been under my nose this entire timethanks anon
okok I know I already beat this prompt to death but just this last one ok
is this a gibson SG?
>>101773928more or lessthe shape is right
Fucking finally...https://files.catbox.moe/hsi2zn.png (embed)Now that we can increase the CFG value on flux, see here:https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/?sort=confidenceIt's time to look at the sweet spot between the Guidance scale and the CFG, and for that we'll use that prompt:>A drawing of Hatsune Miku with dreadlocks and a black skin showing her fists on the streetThe goal here to check whether the photo shows Miku having a black skin + dreadlock, if neither is present, the photo is eliminated.Here's my observations:- CFG = 1 won't change Miku's skin or add any dreadlocks. This is one of the reasons why it's important to have a higher CFG if you want Flux to perform as well as possible in terms of prompt understanding.- The sweet spot seems to be at CFG = 5, that's the minimum CFG where we got the most success.- Having a low guidance doesn't work well with a high CFG, the images gets broken.- High guidance seems to make the model worse at prompt understanding, for example at Guidance = 5.1, there hasn't been a single image that was successful.
Fucking finally...https://files.catbox.moe/hsi2zn.png (embed) (embed)Now that we can increase the CFG value on flux, see here:https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/?sort=confidenceIt's time to look at the sweet spot between the Guidance scale and the CFG, and for that we'll use that prompt:>A drawing of Hatsune Miku with dreadlocks and a black skin showing her fists on the streetThe goal here to check whether the photo shows Miku having a black skin + dreadlock, if neither is present, the photo is eliminated.Here's my observations:- CFG = 1 won't change Miku's skin or add any dreadlocks. This is one of the reasons why it's important to have a higher CFG if you want Flux to perform as well as possible in terms of prompt understanding.- The sweet spot seems to be at CFG = 5, that's the minimum CFG where we got the most success.- Having a low guidance doesn't work well with a high CFG, the images gets broken.- High guidance seems to make the model worse at prompt understanding, for example at Guidance = 5.1, there hasn't been a single image that was successful.
>>101773916np, if you have/get the Impact pack of nodes you can just use this mask detailer sampler, which makes it a bit more simple to setup. Also has face detailer nodes you can use with detection models to auto mask faces/people/hands https://github.com/ltdrdata/ComfyUI-Impact-Pack
>>101773683Late but I thought the pony person was still in the process of making a v7 of the SDXL finetune and was waiting on the Open Model Initiative to put a model out. Flux wasn't planned, it was just a coincidence the creators had it ready last week and marketed it at the right time after the SD3 fiasco. I get he probably wanted to use Flux since it is here right now but he's going to have to wait to make money.
>>101773933very interesting results, high cfg is working nicely it seems (when using half cosine) .. pic related is cfg = 8, guidance = 3.5 .. which seems to be a sweet spotone of my observations to add: high CFG works better for landscape/cityscapes/surrealist/anime/cartoon, for photorealistic human gens I wouldn't go above 5 .. maybe even less
>>101773992Sup
>>101773917I will continue to goon thank you very much
>the preview looks absolutely perfect>oh God, please let it be good, please, please>VAE decode bar is filling, doki doki, waku waku>image is saved>it's awfulI want to cry.
>>101774010>still in the process of making a v7 of the SDXL finetuneno one knows, he had a serious fight with SAI, not sure if that still he still wants to do SDXL, last he blabbered was he wants to make an AuraFlow finetune
and the strat
>>1017740617 string strat.. fancy
>>101774010I hate to break it to you but Pony is a fraud and what you got already was a lightning in a bottle strike where someone incompetent sometimes accidentally makes something good. Pony basically is the YandereDev.
>>101774019>one of my observations to add: high CFG works better for landscape/cityscapes/surrealist/anime/cartoon, for photorealistic human gens I wouldn't go above 5 .. maybe even lessyeah I have the same conclusion, desu having a cfg of 5 is enough to make flux as good as possible on prompt understanding, no need to overkill yeah>high cfg is working nicely it seems (when using half cosine) yeah, using Half Cosine Up on mimic_mode and cfg_mode was the last thing needed to make DynamicThresholding completely viable, glad I found that mf through a XY plothttps://files.catbox.moe/b4hdh0.png
does this flux schnell safetensor run on A1111? Or is some comfy only for now thing?
>>101773508>that hasn't been true for like a year now. You can draw a mask within a load image node and pass the image + mask straight to a sampler.comfy inpainting gradually fries the image to shit if you do a large number of repeated inpaints on your results, as the VAE encodes/decodes the image over and over
>>101774169>A1111?no, two more weeks>comfyyes>>101774155anime gens really like high cfg tho, this is cfg 8, on 5 the sword was borked and the sakura pattern on the kimono was not present
>>101770020Does boomer prompting work because the t5 tokenizer has a max sequence length of 512, whereas the CLIP tokenizer has a max sequence length of 77? Does flux use both or just one or the other at a given time?
>>101774169just use comfyUI bro.its literally better in every way.
>>101774211>anime gens really like high cfg tho, this is cfg 8, on 5 the sword was borked and the sakura pattern on the kimono was not presentoh cool, I'll keep that in mind when making anime gens then
>>101774220by default if you feed a simple prompt into it it uses both with the same prompt, you can use CLIPTextEncodeFlux node tho, then you can send CLIP and T5 each their own prompt, which may or may not be a good idea depending on your prompt and gen subject
What the flux?Is this broken : https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellActs like it's working but then "error".
>>101774274works for me
>>101774211>on 5 the sword was borked and the sakura pattern on the kimono was not presentI guess that on cfg 1 it was even worse?
>>101774274>https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellit was bronked 4me, the dev one in spaces is working but gives like 5 gens
>>101774274
>>101772047recipes?
>>101772047Ah that's why it looks like shit
>>101773627This isn't Flux is it?
>>101774326pic related at cfg 1, didnt get the style at all>simple anime style, flat colors, cell shaded, thick line art, style by Leiji Matsumotoalso the kimono is borked and the swords .. well are suddenly swords
Next bread is ready and waiting...>>101774288>>101774288>>101774288
no collage in the next one, rebake
>>101774058>he had a serious fight with SAIStupid drama or something that makes sense?Didn't even know Pony dude discussed with them, I thought most model makers are silent about what people do to add back porn.
>>101774382every block .5 except the last, final block 1.0 (all Flux)
>>101774468only saw it mentioned on his dev update about pony developed on civitai .. seems there was a discord drama involving license of SAI employees and SD3 license debate, I don't much care, but its the kinda popcorn drama you see so often these days
>>1017744655 API credits has been deposited to your BFL acc
>>101774040redo it and save at earlier steps.
>>101773407Going with the one on the left.On a side note, can Flux do transparent backgrounds? If so, I'd wager it could make an amazing sprite set for Sillytavern.
>>101774580No but it's easy enough to remove a background. It's even built into windows now.
>>101774058I was just wondering if there was any update after https://civitai.com/articles/5069/towards-pony-diffusion-v7 since that was written right when the SD3 debacle started. Any news source for the Auraflow thing or is this Discord talk?>>101774122That's probably not in question, given his past remarks on not knowing things that should be in 101s for anyone doing this semi-seriously. But codemonkeys do occasionally still do good things and he had retard levels amount of money and enough clout and given most finetuners have been hired away from the community and no one else wants to do NSFW tunes, he would be one of the people to watch.
>>101774580Not that I know of
>>101774630The ponyfaggot is sticking to his guns on Pony-AF even though AF is dogshit.
>>101774630>Any news source for the Auraflow thing or is this Discord talk?someone posted a twitter? or other social meda link/screenshot some threads back here
>>101774580there are nodes to accomplish this very easily in comfy"layer diffuse"
>>101774730Okay, I'll dive for it. But AuraFlow does makes sense given his timeframe if he wants to pump something out to keep the moolah going because of the license, but I would think he would try and request early access from the guy and maybe spend money to speed up the creation of the model but apparently, he's just waiting. What a leech.
>>101774463>sees first post Time to add the pastebin already?
>>101774122What kind of resources would it hypothetically take to finetune a flux model? Do we even know?
>>101773826>>101773626catbox?
>>101774964Probably a serious cluster of H100s to do anything like Pony did. But it depends how easy the model is to teach, SDXL was particularly bad.
>>101773407The one in the center is getting her zen on. I like that. I'll go with her for that reason alone.