Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101795805>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
Why do we need two models for clip for flux?
>>101799537One to send telemetry back to BFL
>>101799537Your entire post as the prompt
>>101799537It was trained on both clip and t5 is why
flux is pretty fucking great at img2img/inpainting actually
>>101799609workflow?
>>101799465hello friends. Do you believe me that I have been gooning for at least 5 hours yesterday
halo: AI evolved
>>101799598ok
>>101799621https://files.catbox.moe/jetxy1.png
>>101799640Can you explain what you do? I see you you use two models and not just flux but beyond that I'm not sure I get it.
>>101799669sure, I'll explain.first I create some image normally in a pony based model, just use whatever SDXL format model you like. then I use my self written nsfw recog based node and impact stuff to detect nsfw parts, make masks out of them and invert them.then I just use inpainting with flux, I found the DPMPP_2M_SDE sampler with heun and an eta around 0.01-0.03 best for inpainting/img2img with flux. then I just display the image, the last part is just using the mask to censor like picrel lolI'm a noob at comfyui, I don't know if there are better ways to detect NSFW so I quickly written my own node, maybe but its overkill, but when I tried to load the nsfwrecog model with the ONNXDetectorProvider from Impact Pack it failed
>>101799588Lul
>>101799763I was about to check where the NSFW checker node was but I see now you made it yourself.
>>101799763It (whatever you're spamming) just ends up looking very obviously like SDXL slop thoughJust use SDXL and spare yourself the trouble
>>101800037you're a hateful chud
do you guys like shorks?
here is another one
>>101800057I'd love to try your workflow, but without the NSFW checker going into the segs node it's kind of difficult to use.
Good afternoon
>>101800145hmmm let me make a burner github
>>101800142flux? proompt?
>>101800204NTA but it looks like pony to me.
>my 2 post in opNice
>>101800218thanks for your useless speculation, anon
>>101800057>y-you're a chud!I just don't get what you're trying to do here. SDXL without any other enhancements (not on Comfy even) gives you arguably better results, adheres to your prompt better and doesn't have to go through a million nodes. What's the objective of that workflow again?
>>101800204Using imgnAI, FurXl model. Prompt:>pixel art, smug, shark girl, cute face, teasing, young shark girl, character, blue hair style, petite, full body, short, blue white striped panties, game cg, retro art
>>101800270thanks anon, that style looks great
>>101800234I guess my speculation wasn't so useless after all>>101800270
>>101800322it actually was
>>101800360Actually, it wasn't
anime girl uses comfyui:
>>101799629if I have images like this, I'd goon for 10 hours minimum
>>101800145https://github.com/goburiin/nsfwrecog-comfyuihopefully it works for you, I'm not sure about the requirements.>>101800255>I just don't get what you're trying to do here.obviously not creating gooner AI sloppa, the nunslop is a toy example for an automatic nsfw detection and inpainting workflow with flux. If you don't see any value in that then I don't know what to tell you.
>>101800532>obviously not creating gooner AI sloppa>spammed the exact prompt over and over to get attentionAlso, if you're making a workflow for nsfw detection and inpainting over that, what other applications does that have other than to create coomer slop? lol
>>101800578whatever kill yourself
>>101800585rightShow me a usecase of your workflow for something that is not coomer slop and I'll believe you
>>101800606Hmm well I guess the idea is coomer oriented but that doesn't change the fact that it's an interesting idea
>>101800227Nice
fuck i hate these schizo workflows turn up cfg set negative empty guidance 100 dynamic thresholding number tweaks all for 2x longer gen times
>>101800160Mayhem peruttu
>>101800801but muh negatives tho
now we can do generations with generations within the generations:
>>101800801>>101800969Just don't use it you niggers, are you for real complaining about anon trying out things and sharing his findings?
>>101801006
>>101801204you should consider jumping off a bridge
Flux + NaturalVisionXL makes great photos
>>101801200>types too slowly>dragon roars
>>101801219just report and ignore he craves the attention he lacks IRL
>>101801204I don't care about the kid, how the fuck did you get such a long, coherent text?
been away from this for more than a year.what's the state of local img generation now? can someone give a qrd, and also I'm curious is flux or sd3 worth it? thanks.
>>101801398with the release of flux, local img generation is as good as DALL-E or better in some casesno NSFW yet though
>>101801398SD3 is absolutely NOT worth it. Flux does everything SD3 was supposed to.
>>101801527>SD3 is absolutely NOT worth itSD3 (beta) is not worth it, the 8B model thats API only isnt bad but its not local so who cares
>>101801398Flux without a doubt, Flux mogs sd3 easy
>>101801657>>101801673>>101801686go away fag
>>101801719Idea. If you're brave enough to risk the backlash of a rejected report, you potentially have the power to be rid of him.
>>101801629Prompt please and what model and LoRA did you use?
Wake up jannie
>>101801809let them sleep they deserve the rest
>>101801841We get it. No one is more deserving of a rest than our overworked jannies. All the effort would have gone to waste, until their hour came again.The right janny in the wrong place can make all the difference in the world, so they needed to wake up.
Long live Flux! o/
Finally a Model that gets Swastikas right. NatSoc Propaganda Dream <3
>>101801966Maybe show it to /pol/
I have a 6GB card, and tried to load realvisxlV40_v40LightningBakedvae.safetensors but I can't do anything because I'm out of vram. I tried stetting with PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:25 but even with chunks this small I still can't render anything
>>101801983/pol/ is too stupid to use it
>>101802009Then be their propagandist. Charge them money for your prompt engineering service or something. If you're gonna make nazis, at least get paid.
>>101802002Forgot to mention, I'm not new but just deleted everything and I'm coming back after a while, I remember there was like a 4GB model or that would allow me generate. Is there any other good model I can use?
>>101799582nice
>>101801227
>>101802002>using XL models on vramlet cardsoof
>>101802057could be worse he could be using flux on a 1080 like that one anon god bless him
>>101802057I know, my setup needs an upgrade, was cool 6 years ago when I built it tho
I love abstract art only when it's cyberpunk
>>101801779just Flux
WTF happened to /sdg/
>>101802223Would a 6gb 2060 be enough for FP8? Or would I actually need a 16gb card?
>>101802284/sdg/ is a chat room for avatar fags>Hellooooo cumbot>Hiiii fago, how was your day>Oh you know me, hey has anyone seen dikdik today?
>>101802284It's now a designated containment general for one special individual
Let's say I'm at 74 tokens, would adding a multi-token tag like "looking at viewer" cause issues, since it's split between the two token batches?
>>101802304I think 8gb is that minimum to be able to do anything with Flux, but I haven't really looked into it. Just seen some posts in passing.
>>101801227pretty good
>>101802304It would be extremely painful.
>>101802339cute
>>101802338Yeah probably I doubt there is any intelligent parsing going on
>>101802339Dang, getting a 4070 ti super soon so guess I could wait>>101802353kek
>>101802338yes
>>101802391>>101802401Dang, I hate it. Wish there was an extension that highlighted the prompt according to the token count. Like, you hover the mouse cursor over the token count and it highlights the prompt with different colors according to which token group they belong to.
>>101802057
>>101802505i hate 1girl posters but this gen is cool.
>>101802074>Prompt executed in 716.61 seconds
>>101802535>wait 10 years for a gen>it fucks up the handsmust be painful
>>101802535>handsoh no no fluxanonsit's over.
>>101802535uh, what happened to her arms and hands?
I have yet to harness the full power of sdxl.
>>101802574>>101802585>>101802591everybody is allowed to make mistakes every once in a whilenobody is perfecthow the fuck is she holding the one on the right anyways?
>>101802351Thanks
>>101802647yea its impressive anyway that a 1080ti can even run this at all
>>101802052nice is this flux to sdxl workflow? or viceversa
>>101802734Yeah Flux -> NaturalVisionXL with some inpainting. Just flex (pic rel) is pretty good too
>>101803057nice one anon. I'm interested in the artist and style tags
Okay, it finally feels worthwhile to set up an image generator locally now that flux is out, at least for me. My current experience is exclusively with MidJourney Does Flux work with existing Stable Diffusion UIs? Or is it it's own seperate thing? Would it be easy for me to set up?Also, I remember back when AI was first on the upswing there were rumors that you could fry your GPU by generating locally. I assume that was just retarded fear mongering, right? Or should I actually be careful with my usage?[spoiler]And how good is it with celebrities and fictional characters? As good as peak DALLE?[/spoiler]
>>101803209I don't share due to a thread schizo having a multi year vendetta against me and has no problem impersonating me and threatening to doxx ect
>>101799598>It was trained on both clip and t5 is whydo we know why? they felt like t5 wasn't enough on its own?
>>101803272that just sounds even more schizo anon!
>>101803272>I don't shareif you want to gate-keep prompts feel free to post in >>>/g/sdg
>>101803426>>101803394You must be new here
>>101803272
>>101799763Thanks for the explanation anon, I like the results, I think if flux knew more poses it would be better to start with it.
>>101803272>I'm a victim!tranny behavior
>>101803491i am not, ran.
>>101803597oh, it's the ben10 guy? lmao
Can you make stuff like cartoon style picrel (from dalle), mainly the "wet skin/hair" effect?I know the pose itself flux probably has no idea what it is.
>>101803597>>101803577>>101803426>>101803394Since you're doing pretty bad today here's a pity imagehttps://files.catbox.moe/ee3frg.png
>>101803518>>101803533Anon?
>>101803666deep-fried flux, my favorite
>>101803666that wasn't so hard now was it, ran? next time someone asks for a box give it or leave. PLEASE keep your avatarfag nonsense out of here.
>>101803272>I don't share due to a thread schizo having a multi year vendetta against me and has no problem impersonating me and threatening to doxx ectYou realize that if this is true it's only because you're a faggot that doesn't share, right?
>>101803710it has no metadata in it kek
holy nogen
>>101803733ranfag...
>>101803718said anon has impersonated multiple other anons before, and it's the reason no one shares anymore. but then again you just might be him seething that your thread is dying.
>>101803761>clearly a mentally ill schizo Haha I believe you.
>>101803761>but then again you just might be himmeds can help you with the paranoia, you'll need therapy for everything else though
You are not entitled to a catbox.
>>101803781then why even bother posting the image if nothing can be learned from it?
Thanks for the free bumps to our general
>>101803761>impersonated multiple other anonsthe high iq avoid this issue by not avatarfagging
Give it straight to me, bros, is it possible for a regular person to make money with generative AI? Has anyone here managed to monetize their generations?
>>101803250>Does Flux work with existing Stable Diffusion UIsWorks with comfyui, check OP for linl>Would it be easy for me to set up?If you have basic knowledge about git installing, yes. Otherwise, learn and follow a guide.https://medium.com/@yushantripleseven/installing-comfyui-linux-windows-b59a57af61b6>I assume that was just retarded fear mongeringYes, unless you have a crappy driver or card, never heard of people frying their card in generating stuff.You can cap your card though, if you have a 3090 for example, 275-300W instead of 350W would give you almost the same perf for a big reduction in power usage.>how good is it with celebrities and fictional charactersDoesn't know most characters, sometime you get them back by describing them. The dataset was probably cleaned beforehand.Same for celebs.On that front Dalle is way better, as long as you don't get dogged.
>>101803808It is there for your viewing pleasure, nothing more.
>>101803817>by not avatarfaggingthe problem is that avatarfaggots only manage to get 1 "good" prompt done and then they post endless variations of them, they usually dont have the creativity to keep coming up with new ideas.the real solution would be to have a dedicated AI board on 4chan and a containment thread for these subhumans.
>>101803808have you forgotten how to describe images with words?you don't want to learn, you just want to copy
>>101803848>containment thread for these subhumansWe have one, it's now called /sdg/.
>>101803853>you just want to copywho gives a shit?why does it bother you to begin with?
>>101803848>Same posters can make good work with different models>They continue to excel year after year>names become synonymous with good gensYou sound salty
>>101803865evidentially you because you're bitching people aren't bending over backwards for you
>>101803867>names I cant even name one and I dont care lol
>>101803822do you get multiple you's by people liking your gens/asking for catbox everytime you post?if not reconsider or get better
>>101803867>names become synonymous with good gensexamples?
>>101803877go to the containment thread, you arent welcome here if you dont contribute anything.
>please say me
>>101803822No.
>>101803893>nogenIf you think begging for prompts is contributing you're mentally ill
>>101803893this is chang's thread, if you don't shill for the glorious leader, you don't belong here
>>101803894fuck you that's my line
>>101803907no, being mentally ill is this >>101803272
>>101803924the biggest fear of these faggots is that someone "steals" their prompt and then makes money with it somehow (never going to happen anyway).kek
>>101803924You always get random paranoid anons in these types of threads, just ignore them.
>>1018039111chang has thousand times more creativity than all avatarfags under heaven and earth put together
Lazy prompt beggars are 10x worse than avatarfags and you can't disprove it
>>101803272I've noticed it's generally the mentally ill who always think and claim people are trying to doxx them. Pure schizo behavior.
nogen melty
>>101799465Can I use Stable Diffusion XL with just8 GB of VRAM? I'm using Stable Diffusion 1.5 right now.
/sdg/ refugees ruined this general
>>101803982Which ones?
>>101804003the ones that act really feminine
>>101803648>wet skin/hairFor some reason it doesn't seem to have this concept, maybe "wet" was scrambled when it was in relation to a person/character.
DALL-E 3 still the best, you should all be focusing your brain power in having it leak.
>>101803999checked, also these always turn our beautiful.(ot: old render)
>>101804036>in having it leakthey will never leak this shit
>>101804059usually leaks are against the owner's wishes
>>101804036>focusing your brain power in having it leakIt will never happen anon, even getting an unfiltered access to the API is crazy hard.
>>101803933it's not paranoia, it's annoyance from low effort leechesimagine being given an infinitely creative tool and the first thing you do is type in "Messi in the style of Picasso" and worse, get mad that someone who typed that in doesn't catbox for you.fuck, the whole point of this shit is exploring and all you do is type in 1girl
>>101804036I'd rather see people focus on adding back the good stuff to flux, way more realistic than leaking a model more well guarded than some state secrets.
>>101804090>the whole point of this shit is exploringAnd sharing nice stuff, so that others can improve on it, share back etc.I've learned a lot just by looking at anons posting workflows here.But if you don't want to share, it's up to you, just don't make it some moral statement about it.
>>101804150Notice how he won't post anything. Nobody is here to spoonfeed you schizo
>>101803955Now we're talking
>>101804150you haven't shared shit and I'll be honest, everything here is amateur, like a 14 year old kid showing off his crusty game dev code like he made a masterpiece, you aren't that specialwow you made a comfyui workflow using someone else's plugins! wow that's crazy anon I bet you figured it all yourself and didn't watch a youtube video
>>101804058>>101804098>>101804119>>101804198Gentlemanly
>>101803648>Can you make stuff like cartoon style picrel (from dalle), mainly the "wet skin/hair" effect?I got something like this, what do you think?
>>101803272Frank Frazetta is the artist, Im guessing
>>101804036>DALL-E 3 still the best99.99% of my prompts dont work in DALL-E 3 so there's literally zero proof of this.But I'll take you word for it.
>>101804204suffering from success
>>101804224>>101803648that face is disgusting
>>101799465Flux is puritan garbage.
>>101804270you mean get filtered by GPT or the NSFW check afterwards, DALL-E 3 itself is more than happy to generate smut.
>>101804286>that face is disgustingYeah your're right, I improved it a bit, I think she looks beautiful now
>>101804369unironically much better
>>101804369>not that weirdo who puts tampons in boys bathrooms and calls himself a generalmissed opportunity. still an improvement though!
>>101804406Boys can have periods too, chud.
>Me: "Draw woman carrying a man ">AI: "Here's a man carrying a woman"Are there any non-sexist models out there?
>>101804432>schnell>wow flux doesn't understand my prompt
>>101804432"A woman carrying a much smaller man. The woman is much taller than the man. The man who is being carried is short like a midget."Try that.
Enter!
>>101804487nice. catbox?
you are not entitled to one
>>101804224The idea is here yeah.I've tried getting wet skin so many times, the model doesn't give a shit.Maybe it's the fact that you asked for her to be in the beach.
>>101804487prompt?
Debo needs a wellness check
>>101803648dunno, but try negative: figurine, repeling water, big drop of waterpositive: (realistic skin:0.2)
>>101804514>I've tried getting wet skin so many times, the model doesn't give a shit.>Maybe it's the fact that you asked for her to be in the beach.No, it has more to do with this method that reduces flux's bias greatlyhttps://reddit.com/r/StableDiffusion/comments/1enm9og/discovered_by_accident_a_trick_to_make_flux/
>>101804567is prompt weighting working now? it wasn't before
>>101804592No it doesn't, it's noise.
>>101804406kek'ed
https://huggingface.co/comfyanonymous/Freeway_Animation_Hunyuan_Demo_ComfyUI_Converted
>>101803272haha
>>101804611ni hao
>>101803648or try oil and change hair types
>phones home to CHYNA
>>101804574I'm playing with this, it seems to be able to make her look more "oily/shiny", but at the cost of making the cartoon style go too far.The fun thing being that I can see in latent that at low steps it gets the style right, but at some point it goes caricature instead of cool sexy pixar girl.
>>101804762You should tinker with the GuidanceNegative value, it's at 10, maybe it's too high for what you want to achieve
>>101804689I'll try to add that.>>101804786There was nothing in my test.Basically the cfg trick works, but at the cost of amplifying everything.Wet -> you get that shiny lookPixar style -> way too cartoonySo testing at cfg 2, I get proper style but not wet, and testing at higher cfg, for example 6, I get wet skin but the generation goes too far style wise.Maybe there is a way to amplify only the parts you want with this trick.
>>101804895you're talking about cfg, but did you change the GuidanceNeg value too? maybe that's the one that can help, there's even a third solution, "AdaptiveGuidance threshold", basically it puts cfg = 1 at the very last steps, and you get to choose at what steps it starts to shift from CFG > 1 to CFG = 1https://reddit.com/r/StableDiffusion/comments/1enxcek/improve_the_inference_speed_by_25_at_cfg_1_for/
I just woke up. Did I miss any Flux developments in the last 8 hours?Fuck my weak body needing sleep
>>101804994>Did I miss any Flux developments in the last 8 hours?you can improve the speed of your render if you're using CFG > 1 now >>101804993
>>101804993>>101805004>Download this modified "dynthres_comfyui.py" scriptNo. I am too dumb to know that this does.
>>101805037I wish someone could make a PR on the Dynamic Threshold repo so that the owner can officially implement the modification
I hate catjak now and will dedicate my life to impersonating him sheerly because he wont share his prompts.
>>101804994from the moment I understood the weakness of my flesh...
>>101804611Shouldn't you be called comfypseudonymous?
>>101804993Yeah it's at 10.I thought this was just a speedup thing, I'll try it, too bad it's not modifying the node itself in the git, would be easier to manage.
>>101805081>I thought this was just a speedup thingno, GuidanceNeg is the secret sauce that help flux to stop being biased towards generic slop styles and actually follow your promptshttps://reddit.com/r/StableDiffusion/comments/1enm9og/discovered_by_accident_a_trick_to_make_flux/
catjak here, im gay
dogjak here, Im heterosexual
birdjak here, they're both gay heterosexuals
>>101805205dogjak is out to get me. Help! AAAAAH THE VOICES. Rapeman where are you?!
Debojak here I'm schizo
humanjak here, I'm jaking off
horsejak here. yall got nothing
>>101805142>>101805205>>101805219>>101805227>>101805242>>101805250>>101805252please, call me ranfag
Schizojak here I'm debo
>>101804994>Fuck my weak body needing sleepNo. Gen more.
niggerjak here, im violent
Julienjak here, where the kids at?
sonicjak here, i'm gooning
>>101805054>>101805294
>debo is crying
>>101805328seems like you came to the right thread lmao
>>101805054Is Mechanicus good? I completely forgot about it but that quote just reminded me.
>>101805386nice
Look at the improvement we've made in just a few days: https://files.catbox.moe/njz7qq.pngTo make the workflow work, use this tutorial:https://reddit.com/r/StableDiffusion/comments/1enxcek/improve_the_inference_speed_by_25_at_cfg_1_for/
>>101804993I wouldn't call this a 25% speed improvement, it's just making the last 5 steps quicker.Now, those last 5 steps do go quick. Can you somehow apply this to every step? Then we would be talking about significant improvements.
>>101805642>Can you somehow apply this to every step?if you apply on every step it's the same thing as just having CFG = 1, and yes you can improve the number of steps by decreasing the Adaptive Threshold value
>>101805642Turning CFG over 1 slows down generation time. Anon's script modification starts with a higher CFG, then sets it to 1 once it matters less to the image. You can lower the threshold on the AdaptiveGuider node to switch sooner.
>>101805718>>101805730Alright so the last steps where it is using adaptive threshold are going at normal gen speed, which for me is about 1.19s/it, using dynamic thresholding each step is about 2.30s/it. It's about 12% faster, not 25%.
>>101805855it depends on the pictures, for realistic shit, it can switch to CFG 1 pretty quickly
>>101805718Isn't uncond zero scale > 0 supposed to improve the last steps when cfg is back at 1?
>>101805855>It's about 12% faster, not 25%.it's not just about the speed though, it removes the glitches the high CFG is doing during the begining of the inference steps, look at the second to last and last picture there, the last picture is cleaner and there's no the weird textures you can see on the second to last picture >>101805619
>>101805897Tbh I have no idea what it does kek
>>101805902>it's not just about the speed though, it removes the glitches the high CFGInteresting, there's a specific crispy gen I have I want to test that on
Is he trying to nuke this general
why is it inpossible to generate something far away? Especially without depth of field
>pedophile has evolved into hebephile
>>101806020Do you consider that good or bad?
>>101805962>>101805902Hmm, not that much difference. Needs more testing
>>101806049decrease the adaptive threshold value even more
>>101806059>>101805977Die
Original DT gen always on the left, Adaptive always on the right. Will post more examples.
>>101806092Actually this one doesn't count because the right gen was 20 steps and the left was 50Proper comparison picrel
>>101806092You mean on the left it's CFG > 1 + DynamicThreshold and on the right it's CFG > 1 + DynamicThreshold + Adaptive Threshold?
>>101806130Yes
>>101806127How many steps did it ignore (CFG = 1) on those 50? And what's your Adaptive Threshold value?
>>10180617411. AT is set at your default of 0.994, I am not going to change it for these tests.
>>101806203>I am not going to change it for these tests.you should, if you want something less cooked/burned, you can decrease that value and let more steps at CFG 1
>>101801227are you doing flux first and natvis after? gimme workflow bro
>>101802223pls post workflow
>>101806217It's definitely less cooked from the 2 examples I have done, maybe it can be less fried by changing it but that's not my intention, not yet at least. Once I find my crispiest gen I might start playing with that and see what happens.
>>101806317>It's definitely less cooked from the 2 examples I have done,Nice. AdaptiveGuidance is a really cool node, it makes things faster and fix the overcooking made by the high CFG
>Freeway demo is hereFirst tries. Hunyuan bros we are so backhttps://files.catbox.moe/3fj3n7.pnghttps://files.catbox.moe/sz3rul.png
Extra hard to tell on this one,
>>101800765Yes i can
>>101805619What does 0.993 even mean?
>>101806410pixartfags in shambles
>>101806428its a number
>>101806092Is this overcooked look on purpose?
>>101806436thanks anon.
>>101806410https://files.catbox.moe/l5e4ac.png
>>101806428thredhold = 1 means it won't leave any steps at cfg = 1; the lower this value, the more sensitive it will be to changes, and the more last steps there will be at CFG = 1.Basically in simple terms, if you want less overcook from the high CFG, decrease that adaptive threshold value
>>101806474So it's relative to the number of steps? Wouldn't it be simpler to have a value corresponding to the nth step it changes cfg to 1?
>>101806507that's not how it works, every image is different and have their last steps act differently, AdaptiveGuidance looks at the difference between 2 steps and see if it's bigger than the threshold or not, it's a better way to get something that works for every kind of pictures
>>101806524Oh ok.
>>101799629I'm having a really hard time genning anthro girls, PLEASE share a catbox patrician
>>101806433RIP. Maybe when Bigma comes around it can be this good kekhttps://files.catbox.moe/ex6e8o.pnghttps://files.catbox.moe/on9avi.png
>>101806410I really hope this mf will also finetune Flux now
>>101806443No. It's just what happens when you use a high CFG on flux. The Adaptive Thresholding is intended to run the last steps at CFG 1 like flux is designed to do, but still keep the style adherence of using high CFGs. What I am trying to do is get the same result but less fried.
>>101806611With what money
>>101806648Yeah I've noticed it, it changes everything to the extreme, it's annoying.
>>101806680he wasted that money on huyuan instead of flux, what a retard...
>>101806708MOOOODS
for the anon earlier today
>>101806708can you do a loli but with a huge rack?
>>101806611https://files.catbox.moe/6iz2h6.pnghttps://files.catbox.moe/eg9y9x.pngI'm sure we'll get something for Flux eventually
>>101806763don't feed the troll
>>101806696>he wasted that money on huyuan instead of flux, what a retard...Anon Hunyuan is cheaper to tune, was here first and I'm pretty sure the source of this tune is Chinese. Just thank God we got something that is open more than anything, and Hunyuan is absolutely a great base model I have shown many times. The only reason it's not Dalle/Flux tier is because of prompt following, the dataset used to train it was top tier.
>>101806779>can you do a loli but with a huge rack?unironically probably not. this is a problem with flux being "too smart", it knows what a child should look like too well. but i can try to make some hebes with big tits surehas anyone gotten really big booba on flux, adult or otherwise? what prompt did you use if so?
>>101806829>Hunyuan is absolutely a great base modelHow is it for nsfw? artists? characters? I'm just getting back since I saw Flux and I'm a little lost with all these models.
At AT of 0.950 you stop seeing any differences. 0.980 seems to be the point where it stops following your style so closely and more Flux kicks in. 0.990 seems to be the sweet spot for prompt adherence and deepfrying.
>>101806575Damn that first catbox is cool
>>101806891thanks for the experiment anon, I'll keep that value in mind. How much steps did it ignore at 0.990?
buncha nerds and weirdos
>>101806943oh fuck
>>101799629post the catbox I beg you
>>101806880Artists and characters it knows quite a bit. Styles it knows quite a few, my fav is manga obviously. For nsfw we just got this finetune.Prompt following somewhere between sdxl and flux.I recommend you take a look athttps://imgur.com/a/hunyuandit-0vrZEn0And https://www.shakker.ai/modelinfo/87e2cc2169934523a2ff82fb12e7206b?from=searchThe first are Hunyuan gens I have posted here, or just search archive for "HunyuanDiT" filename. second is finetune model technical info.
>>101799629Put this in the next collage
>>101807128Put this in the next collage
Doesn't really look like MJ but I like that style
>>101806891same prompt and seed with no DT or DT+AT at CFG 1.0 (i.e a normal Flux.1 dev fp8 gen) for the sake of comparison.I have to keep restarting comfy when using DT and not using DT otherwise it just hangs and locks up my PC which is extremely annoying and makes it hard to test.Clearly this works, but there is still improvements to be made to the workflow and/or nodes.Going back to base flux now.>>101806932Flicked over after step 37, so 13/50Also, someone should bake.
>>101807198>Clearly this works, but there is still improvements to be made to the workflow and/or nodes.If I find something else that fix the overcooking I'll post it there and on leddit as usual kek
How is flux at creating ancient chinese interiors? Like, xianxia/wuxia shop/inn/palace/whatever. Should I bother or would Hunyuan or some SDXL model be better suited for this niche?
>>101807265It doesn't really understand the 'ancient' distinction, maybe poor prompting on my part
>>101807310this looks like a japanese antique shop where everything was made in china
>>101807198Left to right, only changing cfg, 1.5,2.0,2.5,3.0
>>101807051I'll take a look, thanks
>>101807265>>101807310
>>101807381
>>101807051Where can you download it?And why is it a "demo"?
Alright we re about to die. Is anyone baking because if not I'm gonna
I have baked and made le collage finna press post real quick, once I have checked the OP is fine
Adaptive Threshold is a good way to keep the style and remove the glitches we got on high CFG, that's coolhttps://imgsli.com/Mjg2MDc4
>>101807392This one is more along the lines of what I'm after, thanks anon! Do you mind sharing your prompt?I had some buzz on shitvitai so tested some of the models on there (rather than randomly download and delete, pic rel), and holy crap the clarify on flux's details is night and day in comparison. Maybe just the models I tried, but kek SDXL-san get your shit together.
>>101807456do it, the guy below you is a larp from /b/
>>101807500put your name back on Teebs
>>101807506go back to your containment niggerlas
>>101807265>HunyuanHunyuan interiors don't look that realistic by default, it would need a tune, here's>这是一间传统的日本茶室(Chashitsu),用于进行茶道仪式。茶室的设计和布置非常讲究,旨在营造一种宁静和谐的氛围。
>>101807425It's a demo because they had a demo up for a while before they released it on Huggingface, Hunyuan by default is .pt so comfy converted it to safetensors you can download it here https://huggingface.co/comfyanonymous/Freeway_Animation_Hunyuan_Demo_ComfyUI_Converted/tree/main
>>101807548>>101807548>>101807548>>101807548
>>101807567Nice, thanks!
