Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101826470>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
>>101828804>>101828804>>101828804
The shall not be ANY brown tigers in this thread.
To be sure if nf4 is better than fp8 we need to know which one of them is the closest to fp16
>>101828830lol
>>101828857fp8 unet? fp16 unet? fp8 T5? fp16 T5?
>>101828830lower that CFG, anon
sdxl & gtx1060 forever...
>>101828873only the unet must be compared, so fp8 unet VS nf4 unet VS fp16 unet, the t5 must remain the same for the 3 images, whether it's fp16 or fp8 is irrelevanthttps://imgsli.com/Mjg2NDA5
>>101828788
>>101828830>white saturationuse AdaptiveGuidance anonhttps://imgsli.com/Mjg2MTI0
>>101828893Sure, how to use it in forge?
Remember, do not the schizo.
>>101828892bottom car not centered fuck
>>101828903>Sure, how to use it in forge?
>>101828903Dynamic Thresholds are in forge by default I think adaptive guidance is part of it.
>>101828885but anon, forge does math all differently from ComfyUI!
>>101827641>>101828858Forgot to say, it's ~17 it/s in comfy for me at cfg 1 without any schizo guidance tricks
I'm gtx1060 6GB mafia but I believe true winners are rtx2060 12GB guys
>>101828922>Dynamic Thresholds are in forge by default I think adaptive guidance is part of it.if adaptive guidance is in it, the problem is that the code change to make adaptive change + dynamic threshold work was only applied on comfyUi (dynthres_comfyui.py), it should be done on forge aswell
>>101828879if its too low it doesnt get the swastikas and jew stars right, if its too high then it gets that high cfg effect but gets the things right.havent really been able to find the sweet spot yet.
>>101828885this is true but i dont know if he ever mentioned what clip the fp8 had. i made the asian girl picture and i noticed that the model i used was his fp8 model from hf but he doesn't mention what clip was used on the readme. he needs to release fp16 dev.
Found a res and step combo that gives me an image every 45 seconds on a 2080ti, with good quality15 steps at 512x768Finally feels worth genning on flux
>>101828924all the comparaisons must be done in forge only of course>>101828950you have 2 solutions, you use adaptive guidance or you mess up with ModelSamplingFlux to mess up with the color saturationhttps://reddit.com/r/StableDiffusion/comments/1epa7yx/control_the_color_saturation_and_contrast_with/
From my tests on a 4080, fp4/nf4 are barely any faster than fp8. This is mostly for vramlets
>>101828962what prompt do you use to get a good ASS?
>>101828946>if adaptive guidance is in itI don't think there is such thing in forge, it's only a ComfyUi node
>>101828979same, if the quality is worse on nf4 I'm going back to fp8, for the moment we don't really know >>101828857
>>101828980Hatsune Miku (very cute) is bending forward and looking back. She is in her panties and leggings, and the image is a closeup of Hatsune Miku's butt; it has a big tattoo which takes up most of her butt cheek that says "works on my machine". Anime-style photo-realistic render in blender.
>>101828979vramlet reporting in!Can my ancient 12gb vram + 16gb system ram rig finally run Flux?
>>101828997how come this:>is bending forward and looking backis first in the prompt but then you told it>and the image is a closeup of Hatsune Miku's buttand so it basically threw away that first part of the prompt.
>>101829004Yep, absolutely
>>101828995Gonna run some xyz comparing the quality. Watch this space.
>>101829021it's pretty easy to do, whoever is the closest to fp16 wins
hype
>>101828892/sdg/ could never
>>101828924>but anon, forge does math all differently from ComfyUI!It's unfortunately true
>>101829012It was too zoomed in on butt before but I wanted more of her body. That ended up working even if it's inconsistent on exactly what it shows.
>>101829004Yes. It doesn't even fill my VRAM with this. Runs toasty though. >>101828637
>>101829063I've heard there was a node that could make ComfyUi's calculations the same as forge, that would be a good way to test the fp16 on ComfyUi and do some comparaisons with fp8 and nf4
>>101829063You are comparing nf4 to fp8. Use the same model.
>>101829063>comparing the nf4 quant to fp8>AdaptiveGuidance enabledholy retardation batman
>>101829088>You are comparing nf4 to fp8. Use the same model.the difference shouldn't be that big, look at nf4 vs fp8 on forge >>101828885 >>101828698
>>101828809Scatman fucking rules
>>101829096>AdaptiveGuidance enabledare you mentally deficient? the threshold is at 1 it means it's deactivated
>>101829106The Hatsune Miku is mine, and it's NF4 vs NP4, not FP8
I'm on a 3060 and the speed boost definitely works for me. It's about double the speed I had before, feels pretty similar to what SDXL speeds were like for me. I hope baking in the fp16 text encoder will make this thing actually just as good as dev fp8 like promised.
>>101829115are you sure it is changing nothing else?just bypass it, retard
>>101829123>I hope baking in the fp16 text encoderthis will eat the speed gains you got if you don't have enough RAM to offload T5
>>101829106not true https://imgsli.com/Mjg2NDA5
3080 laptop, very pleased so far with what I'm seeing going to compare it to the comfy test 832x1280 with cfg negs and all that shitI had to disable xformers to get it to work too
>>101829126>are you sure it is changing nothing else?holy retardation...
>>101829139so no, you're not, you dimwit
>>101829135that one doesn't flip horizontally the character, does it? on Miku it does
>>101828980"Backside", "tracing the contour of her spine", "accentuates her [good adjective] physique", and "her clothes fit her like a second skin" all work pretty well.
>>101829143Rarely seen such a retarded nigger on 4chan, and I've been here for years:1) Dynamic Threshold = 1 means that it's fucking deactivated2) Even if it's not deactivated, and it does something, all it does is to change the CFG to 1, guess what retard? it's already in 1 so nothing is changingYOU FUCKING NIGGER
>>101829132>this will eat the speed gains you got if you don't have enough RAM to offload T5no worries, I have a gorillion gigs of ram
>>101829159>and it does something, all it does is to change the CFG to 1only a truly mentally deficient baboon would say this with such confidenceyou can't even understand the code of the node, can you?
>>101829148thanks
>>101829063what are the Set Device nodes from? can it force T5 to CPU and run it there?
>>101829171https://github.com/asagi4/ComfyUI-Adaptive-Guidance>The node calculates the cosine similarity between the u-net's conditional and unconditional output ("positive" and "negative" prompts) and once the similarity crosses the specified threshold, it sets CFG to 1.0, effectively skipping negative prompt calculations and speeding up inference.Now I'm gonna ask you to kill yourself
>>101829136It's not even close, I just have Async swap on my forge install and it's running circles around my comfy install. This is the typical time I got using comfy.I do have 64gb of ram so I'm not sure if that's a factor
>>101829184You just keep showing your retardation, quoting docs instead of linking to code to show you understand what it is doingkeep on living in idiocy, codelet
>>101829147flip the character? huh? i have no idea what you are talking about. it's two gens done with the same settings and only models changed. they should be near the same picture.
>>101829186Also --cuda-malloc --disable-xformerson for forge arguments
Posted this in wrong thread because I am a retard.https://imgsli.com/Mjg2NDEyThe difference between fp8 and np4 is minor, and in some cases (at least with this example) better.
>>101829204Why --disable-xformers?
>>101829216There's a bug where it won't start for some systems I'm one of the unlucky ones. I might see speed gains if that's fixed
So is forge better at flux than comfy for vramlets?
>>101829191like I said, I asked you to kill yourself, you're wasting everyone's time with your 2 digit IQ takeshttps://files.catbox.moe/kuedol.png
>>101829216You know why
>>101829215Nice, thanks for the effort. Can't see much of a reason not to run NP4
>>101829234My laptop 3080 has 16gb of vram and it's a night and day difference for me
>>101829198that's what I said, look at this, miku's got horizontally flipped, it can't be just the difference between nf4 and fp8, it's just that forge calculate shit differently than comfyUi >>101829063
>>101829215prompt and settings?
>>101829235>being this retardeddid you skip Logic 101 in school? Showing there is no difference doesn't retroactively make you correct in thinking there was no difference.
Optimization and quantization are great.Nvidia's stock price need to return to where it was a year ago.
>>101829242I don't know if it's just forge, but my GPU gets up to 75 degrees. It doesn't do that on comfy. >>101829252Prompt: Hatsune MikuSeeds: 0,1,2,3,4Euler Beta. 20 steps. 1024x1024. Distilled CFG 3.5, CFG 1.
>>101829245As in it's much faster on forge? Or better quality?
>>101829246i didnt make that picture on comfy i made both on forge. that's not relevant to the picture i linked.
>>101829282It's magnitudes faster on my system
>>101829256Wait a second, you're still alive?
>>101828788What the fuck is Forge, why does it look like Auto1111, and can I use nodes with it?
>>101829294SNIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIF
In case you're wondering comfy is seething
>>101829295Thanks anon, I will give it a try>>101829300Faster A1111 rewrite that was abandoned a while ago but now apparently had a smashing comeback
>>101829282>Or better quality?we don't know, we need a fp16 output on forge to see which one between fp8 and nf4 gets the closest to it
>>101829296retard anonie with his anime images
Do xformers make flux gens faster?
>>101829317go fuck yourself nigger
>>101829327/pol/ is that way
>>101829307why would he seethe instead of just implementing nf4(non-rhetorical question)any answer that reduces to "because he is a bad programmer and does not know how" will be considered a silly shitpost and ignored btw
>check ublock origin in forge>it blocked cloudflare connections for mewhy are there cloudflare connections in forge bros?
>>101828608how many steps anon? I need that to try it out on comfyUi
>>101829186godawful titlebar font neck yourself faggot
>>101829342He didn't care to integrate it and now says he's going to copy it in a custom node to dodge the ToS
>>101829360forge guy is kinda a cuck for saying "no you can't use my optimization" desu
>>101829360nta, why does Comfy need to worry about ToS if Forge doesn't? confusing
>>101829370Comfy and Ani tried to cancel him
https://files.catbox.moe/5uus2r.jpgBeen trying to use Flux for body pillow designs
>>101829360uhh is Forge trying to patent the concept of nf4 or something? lolno one needs his permission to write their own implementation, he didn't invent nf4
>>101829386tic tac nipples...
>>101829370He's only prohibits commercial use which comfy is trying to do with his ui
>>10182935520 steps it will be different on comfy.>>101829280im not getting the same results as you. are you sure you labeled right?
Bros, I don't know who this Forge guy is, but I think we should start an harassment campaign against him to force him to give his tech to Comfy. Comfy is /ourguy/. I bet that Forge guy doesn't even post here.
>>101829370That's not what cuck means desu
>>101829413obvious false flag from a comfy hater, make it more believable next time
>>101829413Comfy tried to ruin him for using free and open comfy code
>>101829412>20 steps it will be different on comfy.can you make a screen on the workflow for fp8 flux + fp8 t5 on forge and on comfyUi? a bit like this >>101829235
>>101829396I tried again and it looks like a more complete body pillow designhttps://files.catbox.moe/mjg8kg.jpg
I sleep
its been too long since we've had a UI war
>>101829422I love Comfy, bro. I used to be a Voldy shill and used to harass anyone who shilled for Comfy here, but, in the end, his ui won me over. The nodes lend themselves to so much experimentation. This Forge guy should give us his tech. I won't move to an ui that doesn't use nodes ever again.
Yeah, I'm at the point where forge faggot needs to take this down until comfy anon gets a look at it. This is bullshit.
>>101829425I'm sure Comfy had good reasons, but even if he didn't, I'm still supporting him, because I like his ui the best.
can people explain why they believe comfy would need to use any Forge code to implement nf4you don't genuinely believe that illy invented nf4, do you
Humanity must stop climate change and must be prepared for giant meteorite impacts and space wars.We need AGI, robots, fusion reactors, and cheap GPUs.And gtx1060 mafias needs to win.
>>101829412
>>101829488>can people explain why they believe comfy would need to use any Forge code to implement nf4because he never implemented it before? he had plenty of time, without forge forcing him to do it, he would've never done it
explain how forge works without downloading text encoder models, just the nf4 worksalso, IS nf4 superior to fp8?
>>101829503>he hasn't implemented it before>therefore he would need to steal code to do it????I don't think this logic holds up
>harass a dev for using your code >code which is free>dev even credits you>dev abandons project>keep talking shit>piss him off to the point he makes a new backend without your code>btfo's your backend Ironic>>101829488Nobody is saying that comfy had no interest in integrating it and is seething it's drawing attention away from his project
>>101829527I never said he would steal the code, what I said is that he wouldn't have CONSIDERED it if forge wouldn't do it aswell
>>101829506yeah, nf4 magically encodes the textyou fucking retard
>>101829488I don't know, bro. I'm not code savvy. But I'm with Comfy here. If Comfy needs this guy tech, then you just have to tell me and I'll start a harassment campaign against him.
>>101829449>UI War
>>101829506>explain how forge works without downloading text encoder modelthe file you're downloading for forge has everything in it, the VAE, the text encoder and flux
>>101829538I didn't have to download the two clip files (clip_l and t5xxl_fp16.safetensors)>>101829553is nf4 better than fp8 or too close to be notable?
>>101829553That sounds so fast and efficient, bro. I can't believe this Forge guy doesn't want us to have this for ourselves. What happened to being open source, bros?
>>101829536oh right yeah, seems uncontroversial that if he implements it now it will be as a response to Forge doing it and therefore creating user demand for itThat doesn't seem embarrassing to me in any way though
Comfy trying to figure out how Illy stole his NF4 code
>>101829500i get this gen on seed 2 fp8. unless we using diff fp8?
>>101829543again you're laying it on too thick, dial it down like 3 notches if you want people to buy it
Any tips for getting bottomless characters? Most prompts I can think of still have panties 90% of the time.
just push the fucking code so i can keep the wires
>>101829571what's embarassing is that he knew he could implement nf4 to make everyone's life easier (especially the vramlet) and didn't consider it to the slightlest, seems like Comfy likes to shit on its users, whethers it's with its shit Ui or by not making implementations that actually matter
>>101829569>I didn't have to download the two clip files (clip_l and t5xxl_fp16.safetensors)you did, just not separately, they are in the file
>>101829569>is nf4 better than fp8 or too close to be notable?he claims it's better than fp8 but we'll see, hard to tell since we're forced to use the shittier text encoder for now
>>101829586even nude or naked doesn't workI feel your pain anon
>>101829591>hard to tell since we're forced to use the shittier text encoder for nowit's not the problem, you can use the same text encoder (whether it's shit or not) for both fp8 and nf4 and if theorically they're the same then you'd get the same picture
>>101829569>is nf4 better than fp8That's the million dollar question that people are currently sperging out about ITT
>>101829573
>>101829590how is it so much smaller? flux fp16 is like 23gb, and the clip models are almost 10gb, while flux nf4 is 11 gigs, even if you get fp8 flux its still larger than this one file without the text encoders
>>101829580Oh I fucked up and missed seed #1 on that link. That's actually seed 3. Here some more:Prompt: Hatsune MikuSeeds: 0,1,2,3,4,5Euler Beta. 20 steps. 1024x1024. Distilled CFG 3.5, CFG 1.Image comparison: https://imgsli.com/Mjg2NDEySimple text: https://imgsli.com/Mjg2NDE0Complex text: https://imgsli.com/Mjg2NDE2Complex Prompt: https://imgsli.com/Mjg2NDE3
>>101829602the only way to know if to get a fp16 file, but we don't, so...
>>101829588What the fuck? Is this real, bros? I thought Comfy's call to fame was that he was all about supporting the users unlike Voldy. Wow, bros. This is not a good luck. But his ui is still the best.
unprompted heart shaped pasties are fucking cancer
>>101829235
>>101829609>how is it so much smaller?because it has quantized some of the values down to 4 bits, it's in the fucking name.
>>101829627Meds.
>>101829622Just use your eyes, it is clearly better
>>101829627
>>101829599not necessarily, one could be different since it's obviously not an identical model, but still retain the same level of intelligence and qualityand the claim is that nf4 is actually better than fp8
>>101829639right, but I assumed there would be a notable quality drop but that doesn't seem to be the case
>>101829610>Complex Prompt: https://imgsli.com/Mjg2NDE3ok I think the debate is over, nf4 is really better than fp8... unless the "complex prompt" was "4 arms" then... :^)
>>101829636see, not the same thing at all, on forge miku is waving her left hand, on comfyUi she's waving her right hand
>>101829645Very erotic
breaking news tv prompts seem to work fine, it's not always perfect but this one works:
I'm getting black images, using this model https://huggingface.co/lllyasviel/flux1-dev-bnb-nf4 with assumingly the right settings i have no idea what i could be doing wrong. Can't believe Forge finally updated, which means i'm VERY behind on this news.>yes this shit took 2 minutes please be nice to my gtx 1080 and potentially wrong settings
>>101829650NOOO WE NEED MUH FP16 TO KNOW IF 2 ARMS ARE BETTER THAN 4
>>101829645When are we getting 3D printers that print our SD waifus?
>>101829650Prompt:>A still frame from a movie starring Hatsune Miku, wearing a black crop top with the print "Sugoi!". She has a beautiful face with black goth smokey eyes eyeliner and mascara, black lipstick, large breasts. her navel is visible. she is wearing a thin o-ring choker and ripped jeans hotpants with a string thong underneath. her arms are covered in sleeve tattoos. She is lifting or pulling her top, the background is dynamic and vibrant.It's only complex compared to the previous ones, where it was just "Hatsune Miku" or "Hatsune Miku saying: "
>>101829667Show us your settings up above
I'm not even using xformers and it's faster
>>101829669If nf4 can be optmised this good, why can't we do the same thing for a nf8 for example? that would be something like fp16 > nf8 > nf4 > fp8
>>101829610heres seed 3 it doesn't match either.i just want to match the broken arm fp8 you posted.
>>101829645How did you make her body black and her face white?
>>101829684
>>101829650Oh fuck, now that I see that nf4 gives better quality, I really want to test it out with fp16 t5 now...
Can confirm, on 12 GB forge is under 20 s/it whereas comfy is over 30 s/itAnd I can probably have it use more VRAM as well as I just limited it around 9
Reminder nf4 works on xl too
>>101829689I don't see why not, the LLM community makes a million different quants for every model.
>>101829660I am using a different checkpoint to you I think. I am using Comfys mix for that. Where did you download yours from?
I set guidance to 20 and I stopped getting hands covering boobs, extra belly buttons, and heart shaped pastieshttps://files.catbox.moe/1m7gfp.jpg
>>101829646idk the details but you can accomplish a lot of compression by doing shit like keeping very important weights at fp16 and unimportant ones way down to 4 bits or less
>>101829697Try "auto" and "queue"Then turn on "shared" after if everything works fine for more speed
>>101829710yeah that's basically that
>>101829694Screenshot the one you mean, I will post catbox
Make great text encoder for sdxl
>>101829610>Complex Prompt: https://imgsli.com/Mjg2NDE3the difference is huge, I really think nf4 is really close to fp16, didn't expect to perform this well, goddamn...
>>101829732>a Trump doppelgangerI fucking knew it
>>101829711still a black image.
>>101829734OK but why is it so good? It doesn't really make sense to me.
>>101829744What does the console say?
>>101829727seed 2 this linkhttps://imgsli.com/Mjg2NDEy
>>101829708Positive or negative guidance?Also those nipples are still fucked bro
>>101829758I don't even have a proper local setup yet. Positive guidance only for now. rip
>>101829755https://files.catbox.moe/xw4fdb.png
Comfy decided to not implement it for the sake of his former coworkers. They wanted to push a SaaS model and this hurts this greatly
>>101829750heres my console m8by the way i got a potential hint as to what the problem could be in the description of this model on HFhttps://huggingface.co/twodgirl/Flux-dev-optimum-quant-qfloat8>In int4 mode there are places where the pre-trained weights in fp16 overflow, resulting in a blank image.still dont know what any of this shit means but i thought i'd bring it up.
>>101829744>1080
>>101829768He didn't implement it because he's lazy and incompetent
>>101829706RIP TheBloke
>>101829769Are you using the copy of the model linked on the forge announcement page? A bad model is the only thing that comes to mind at the moment.
>>101829745it's a mix of precisions, fp8 is (basically) all fp8, nf4 picks and chooses the precision per parameter
>>101829784no, linked here >>101829667what copy of the nf4 quant is known good? I just grabbed whatever i could find on HF.
>>101829788Magic. Got it.
>>101829744You do have a 3000 series card, right?
>>101829790https://huggingface.co/lllyasviel/flux1-dev-bnb-nf4
>>101829768Comfy said in the last thread it would be out tonight. Then again, that comfy didn't post a fennec with the post.
>>101829667nf4 is for 3000/4000 cards I believe, fp8 for all cards
>>101829790That's the right model. I'd try redownloading it I guess. Did you setup forge just now to start testing this new model? Or is it an old setup?
Imagine if you had your avatar on right now
>>101829793It is magic but local LLM chads have been doing it for a while now
>>101829610Wow. This is so good, bros. Comfy bros, I think we may be in trouble here. This guy will need to give us his tech or we'll lack behind. What if we ask politely, bros?
Will we be able to finetune fp4 models?
>>101829794>>101829798do you niggers even read? kek i don't mean to be rude but damn>>101829667>>101829808I'll redownload, could be a corrupt DL since i was on my VPN at the time anyway. I just updated forge to use this, Should i honestly just redownload forge at this point? That might be the more likely reason this is fucking up.
>>101829779this guy huggingface.co/mradermacher seems to pump out as much if not more quants as quickly than TheBloke did.
>>101829805I'm using the nf4 model on a 2080 ti just fine
>it's been 3 years and people are still fagging over a fucking UI
>>101829820I'd try scrapping your old forge and re-cloning from github as a last resort. I had problems with using an old version I updated from.
>angery because anons that got gatekept from using the latest model can now use it
>>101829832>One AI suddenly got non a gimmicky feature that blows the other UIs out of the water>WHY ARE PEOPLE STILL DISCUSSING UIS??!?
FLUX is fubar, I'm not even trying to gen coom shit but whenever it decides a shirt or suit is too tight and breast shape too visible it makes the shirt texture blurry and stretched as fuck.
>>101829832>3 yearsSD came out two years ago this month
>>101829832people should just use whatever they like, and multiple UIs if they like. I'll use forge or comfy for flux, and auto1111 for my 1.5/SDXL/pony checkpoints cause I have a ton of character/style loras.
>>101829766at first i suspected you might be shilling but i imported that png and i get my same picture. i dont think swap location or method should matter. i dont think whatever optimization newer cards get matter but who knows. but they just dont come out the same. can another anon import and gen that pic? and post what it gens? one of our setups is weird or theres a bug.
>>101829848It's not that, it's the fact that there's 1 guy who is such a fucking schizo with a hate boner for comfy that ANYTHING that comfy can't do he immediately becomes a massive fuckwit for hours until comfy pushes an update then he finds some other shit to be mental about. I genuinely hate him and hope literally dies. I don't give a shit which UI someone uses but this guy is fucked in the head about it to an unhealthy degree.
python is only using 7 gigs of memory now while generating. and 9.2 gigs of VRAM. so more systems should be able to use flux, which is good.
is Forge using CPU noise?
>>101829838Thanks bud i fixed it, just needed to do some ((pulling)) in github desktop. Shouldve honestly just updated from there to begin with but i forget i even have it installed.Aaaand this first test was shit. To be fair this was 768x so i would guess its standard res is minimum 1024xhttps://files.catbox.moe/m8pdjg.png
>>101829877Yeah weirdly I get this nowI don't know what happened before, I must have fucked something up somewhere and not noticed
>>101829893*also I had a stream open so it probably uses even less video memory.
>>101829877oops i forgot picturehttps://files.catbox.moe/ph8hw1.png
imagine how many boomers can be tricked with flux
>>101829891Hi debo you misused that word again
>>101829899Minimum res is like 48x48
>>101829919debo doesn't write like that at all
>>101829891Just ignore him. I like all UIs but I use comfy 90% of the time because it gets updates the fasted and I'm not some reddit shitter who's afraid of le spaghetti. But today forge come out with something cool.
>>101829918yeah instead of fighting these faggy UI wars you people could be collaborating on some solid boomer PSYOPS. Get some shit spinning about political candidates and officials meeting aliens and anime characters IRL. We should see this shit on mainstream news in the morning.>>101829924good miku lord.
>>101829877Here, recreated. Here's all my settings.
>>101829932I treat these UIs like apps, you aren't limited to one app. Use them all. Like, flux inpainting workflow is literally dragging one image away and easy, so there is no reason you HAVE to use forge, both are fine.
>>101829919I'm not debo you fuckwit.
>Controlnet + pony to flux with flux canny control net randomly flips lowvram mode without warning.Cumpy can't come soon enough with that fp4 model.
miku 2024
>>101829928>t: reddit shitter who enjoys spaghetti
>>101829949so were all the comparisons done with the beta scheduler? that scheduler can schizo out sometimes. how does it look on nf4 with the simple scheduler?
>spend the entire night prompting againsince flux came out I didnt sleep more than 6 hours.
>he's making it obious againkek
>>101829979How are things going over SAI?
>>101829983I really like flux but it fucking needs 24gb card LoRA training support soon.
>>101829981Yeah all wit beta, it's how I had it on comfy, just copied my settings. Re-did the 4 arm woman on Simple and it looks better.
>>101829983nice honorary aryan
>Flux knows ReimuNeat, though it seems to be confusing her with Flandere just a little bit.>>101829899I gen at sub 1024 fine. Even below 512 you start getting artifacts/blurriness but it tends to be fine still. Also, leave normal CFG at 1 if you aren't.
>>101829997idk why you asked me, but probably not good since its SAI kek
>>101830011could you try the complex text prompt again on both nf4 and fp8 using simple? i think beta nukes texts more than the fp8 t5
>>101829981>>101830011>>101830031In fact here it is compared to the NF4I need to re-do all my tests againhttps://imgsli.com/Mjg2NDIw
>>101830014Flux also knows Vanellope Von Schweetz from Wreck-It Ralph and Fiolina Germi from Metal Slug
>>101829902>>101829909>>101829949 very strange are comparisons reliable then? maybe with simple?>>101829981simple nf4 seed 3https://files.catbox.moe/ppxudj.png
Sirs please do the needful
>>101830039there should be a prompt database somewhere with known characters, loras are possible but still, even with pony models you can get good results with simple booru tags for characters.
>>101830042Still different to yours. Used png info and hit run.Weird
Alright you idiots, I know you shill here. I'll drop 5K/month for 3 months in the pateron of whoever decides to do an anime finetune of flux.
reminder, it's the law
Lame, nf4 flux in Forge isn't any faster on my 3090 than fp8 in Comfy was. I guess only vramlets get a speedup.
>>101830042>>101830061There must be GPU differences. I am on a 4080, what are you running?
>>101830042>are comparisons reliable then?yes, but for the beta scheduler, now we need to check simple. though it will be an interesting result if these models favor different schedulers.
>>101830039>Fiolina Germi from Metal Slughow does it know this character and not more well known ones? can we guess the dataset based on the characters known?
>>101830065even better:
>>101830042this is me>>101830072i have 3080 10gb which i think should be new enough to not have weird issues.>>101830081i check simple with that other anon still different.
>>101830014Damn, okay. What CFG should i be using at say, 768x and 10-20 samples? Should i be keeping samples at 20?feels so weird, having to learn an all new imagegen model after feeling like the drought was never ending..Also i think i saw in one of the threads back, someone posted this particular code to get nudity?Damn just how fresh is all this??
>>101830100it depends on the noise right? cpu/gpu? what does forge use for that?
>>101830107kek
are you guys using forge on CUDA 12.4 or 12.1? 12.4 doesn't work for me for some reason.
>>101830088Can you just like not post pedo shit? It wasn't funny hte first time and it isn't funny now.
>>101830109As far as I know normal CFG should always be at 1, and you should only change distill CFG. Maybe that's changed though. As for nudity, I've found it works decently until you mention clothes at all. If you just say "so and so is completely nude" it'll work most of the time, but if you say "so and so has no panties" it will gen panties.
>>101830125I have cuda 11.8 and it's fine in the various ui's
>>101830133Ignore and report the thread schizo
>>101830103>>101830088Can you just stop? It's really not okay.
>>101830123They are ready to learn from the master.
sometimes all the text works out nice even in challenging prompts (news overlays):
how do i even know if my gpu supports nf4?some 20xx do, according to >>101829823does 2070 super?
>>101830094it probably knows the LoRa model trigger wordsI put "Fiodef" from the Civitai LoRa model trigger word and it works
>>101830152Only one way to find out, vramlet
>>101830133it's the schizo from sdg, their thread is dying, and he needs the attention. he is jealous, just ignore him.
>>101830133Pedophilia is a transgression fetish, he is doing it intentionally to disgust people and piss them off. Having an audience is part of the paraphilia, it's why you see so many stories about them getting caught because they couldn't quietly indulge in it on their own and had to be part of some retarded online pedo community
>>101830152I'm the anon you quoted, I just tested it and it worked, so I kept using it. May as well do the same.
>>101830153wait wtf, flux trained on ai loras?? kek can we just call civitai loras into flux? big if true
>>101830153quick link the lora, once i get forge working i have some science to do
>>101830110im going to bed it's too late. i have a 10850k but the whole point of pnginfo and importing is reproducibility. he genned a strange armed fp8 and then later with the same settings a normal one while i always genned a diff normal one. maybe was beta at that time but he shouldn't be able to get two diff gens huge quality gap too, either way i don't trust comparisons now, i think something is wrong.
>>101830166what's your s/it?
>>101830173>flux trained on ai lorasHuge if true
Strange choice of waifu...
>>101830187~6s/it at 768x1024~8s/it at 1024x1024~2s/it at 512x768
>>101830245yeah, i guess it doesn't get better than ~12s/it for x1024 on 8gb vram
I can't show this thread to family because of this guy.
>>101830066but the quality seems higher on nf4, and the vram requirement has been halved, which is always good >>101829610
>>101829642Fucking kek, someone remake this with flux
>>101830270You get competent images at 5 its, and great ones at 10 if you've been leaving it at the default 20.
Lots of discussion on the new Forge update making this thread pretty quick so here's the next one...>>101830320>>101830320>>101830320
>>101830314sorry to hear you're americanhave an old lady in this trying time
