Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101842860>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/ComfyUIForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: 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>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
Is there a way to make the both of them work at the same time?
>>101847080Do you use fotor for the collages?
>>101847097They have the same output. Is there a node that accepts multiple guidance inputs?
Philosopher's Stone era Emma Watson lora where?
>>101847121In your gpu waiting to be unlocked you sick fuck.I don't think anyone is going to publicly make that lora.
Blessed thread of frenship
>>101847144not for NSFW, you're sick in the head>>101847164go fuck yourself
>>101847144>>101847171Calm down, faggotGo be mad somewhere else
>>101847177hes not being mad, he's being fun
>>101847121We're definitely getting an age slider lora that works both ways like it was with sd1.5 and sdxl
>>101847171>not for nsfw
Do we know if finetunes can be made on consumer hardware, or if it's just Loras?
>>101847227nice one anon, prompt and settings?
anybody have lewd chat gpt?
>>101847097Whats that for?
>>101847235https://files.catbox.moe/n7mpx2.pngI've been asking chatgpt to describe the artstyles of images I like, then getting it to generate a T5 prompt to generate an image in that style
>>101847227Very unlikely looking at this stage but this time a week ago we thought any training was literally impossible so who knows what will happen next week.
>>101847252I'm testing some shit to limit the burn of the CFG, but I like adaptiveGuidance and I'd want to keep itCFGLimiterGuider is this shit btw: https://arxiv.org/pdf/2404.07724
>>101847275I see
>>101847275
>>101847275this nigga taking notes, more effort than 99.99% of image gen users
This is what I love about image and text gen. When things get stale you can just take a few months off and suddenly ten years have passed
with the latest update to forge, i'm unable to use PonyRealism, recreating the same seed gives me a 2.5d image. wtf. anyone have ideas? settings seem identical.
>>101847413the calculations, anon, they...different
>>101847377>mouse on left sidedoes miku constantly flick the bean, do ya reckon?
>>101847421no no NO NO NO EVERYTHING WAS PERFECT BEFORE PLEASE DON'T DO THIS TO ME PLEASE SOMEONE DELIVER ME FROM UPDOOOTER EVIL
Oh shit, Euler CFG++ actually works on flux, you just need this nodehttps://github.com/pamparamm/ComfyUI-ppmhttps://imgsli.com/Mjg2NzE2
>>101847391kek>>101847426don't worry, me neither, I just test out every single node that has "Guidance" in it, maybe with some luck I'll find something good kek
>>101847454it doesn't work well with the current values of DynamicThresholding though at high CFG
>>101847492
Oh wait, do the ((new calculations)) have to do with all those new features like being able to choose a specific scheduler and distilled cfg? Do they make a huge impact on PonyXL? I never used them because obviously forge didnt have them before.Shit how would i even find out which scheduler it was picking before? Do the specific samplers have a preferred scheduler? For example im trying to make these peach gens with DPM++ 3M SDE..
>>101847492>>101847454Is there some documentation about that "cfg_++" or a paper or something?
>>101847532you're just fucked, man, time to test settings to find the sweet spot again
>>101847557no fucking way i have so many god damn good gens i liked to use as foundations i can't just start over again
I noticed that sometimes the negative prompt node takes up to 20 seconds to run (while obviously doing nothing useful), apparently because of t5xxl, here's a hack to speed things up.I hope we get a ksampler without a mandatory negative input.
>>101847431Just install the older version
>>101847574>that picturelmaoooooo
>>101847574>dall-e gen
>>101847575>I hope we get a ksampler without a mandatory negative input.that's what SamplerCustom is for, the ComfyUI Flux example workflow uses it. go check it out
>>101847574why not just downgrade?
>>101847621No real point if Flux is the next big thing. Not counting my chickens on getting FluxPony this year but it'll happen.and no fucking way am i bloating my C: with another 5 gigs worth of python dependencies to run two forges. Oh well, dems the fucking breaks i guess ill have to spend the rest of my morning getting good again.>>101847599that's not a gen that's just a picture of Lennon and his wife.
>>101847636>and no fucking way am i bloating my C: with another 5 gigs worth of python dependencies to run two forges...you cant dynamically link two forge instances to a single package of dependencies?
>>101847574one has to imagine anon happy
>>101847377>tattoos
>>101847674Even if i knew you can do that and how, Wouldn't that beat the purpose of using different instances? Pretty sure new forge has new dependencies.
>>101847611Oh nice, I saw SamplerCustom but it had negative input too, SamplerCustomAdvanced is exactly what I wanted.
>>101847693surely you could just have a single folder loaded with dependencies, and have programs link to that directory. right? SURELY coders arent so fucking retarded that the concept of libraries makes them sperg out, right!!!
>>101847687Anon can't even handle ai-generated women
Even LoRAs are looking a little slopped to me. I hope it's just because he didn't take enough steps, otherwise a finetune may be needed.
>still plain miku gensI thought this general would be the more creative one. just a circlejerk for the most updooted meme char
made my first flux lora. It's cool i guess, but the real thing is how little it takes to train! 2700 steps, 22 images, and only one 3090. https://civitai.com/models/640342/shamiko-or-machikado-mazoku-fluxd
the power of a text encoder:
>>101847791No way the first loras that come out are sloppy!?I can't believe it. I won't believe it
>>101847872>>101847791He's just didn't train it right. Mine turned out fine.See here -> >>101847824
>>101847824You wrote 3x3090 in the description, did it take 3 hours on a single one?
>>101847872Not sloppy, I mean AI slopped, which is baked into the base model.
>>101847828You are worse than the hunyan shills.
>>101847886That is also what that anon meant.
>>101847887im just pointing out that it's nice that a model can do text well now
>>101847879I have 3x3090's so there was three versions of the lora made, with different params.Distributed training is broken at the moment for fp8, so I just three runs in three separate gpu's.
wonder if there's still gonna be retards staying on 1.5 like there was when SDXL came out, or is flux so much better that they can't fool themselves anymore
If I'm getting fucked hands should I increase or decrease CFG?
Sirs cant get it to generate this pose help .
is there a good comfy workflow for loras that 100% works with the converted comfy lora samples
>>101847887>worse than the hunyan shills.I'm not shilling Hunyuan. I will simply not stop using it. It takes knowing what SOTA closed aesthetic models like Niji output to know that Hunyuan is the most aesthetic model rn even with Flux being the best at prompt following.
>>101847941technically increase cfg would be better but your best bet is to just regen with 5000 different seeds and hope you get a better one
>>101847941I don't think CFG is fucking the hands, it fucks the color saturation and contrast but the anatomy is always fine
>>101847824Could you please show me some pictures from the dataset and txt used for them, do you need to write a fluid description of the picture in the txt now? I wonder if the old simple tags from my dataset for SDXL training will work with flux.
>>101847968>>101847967Yeah, I think Flux just struggles with dynamic poses or something, it's the first time I've run into a bad anatomy issue.
>>101847454>Euler CFG++>++What's that?
>>101847994One of the runs I did was with danbooru tags, but I also set the learning rate waaay to low so the danbooru tag part of the experiment got shafted. I was going to try it again with danbooru tags sometime.As for what I used, it's a danbooru tag based full description of the image. I've provided a refined example from my current project.I feed the image, and the danbooru tags to either GPT-4o or Claude and have them write captions. It works good enough.
>>101847080very impressive 2B!>>101847824thanks! love thad adorable retardon a side note, is pony supposed to look washed out without style loras? I am using the sdxl vae. Is there a magic word I'm mossing? (Euler A)
>>101848094Thanks!
>4 panel manga, can specify what happens in first/second/third/fourth panelsneat
>>101848108get the autismmix checkpoint off civitai, also the confetti variation of it, does really good anime stuffif you think it's washed out make sure you have a new copy of the sdxl vae. but appearance varies by model.
Oh fuck, I got some new schedulers after downloading dozens of nodes that night kek
Any advice on a model/prompting to generate something without people? I'm trying to generate an empty wrestling ring, no crowd, from the point of view of inside of the ring. The more I ask for no people, the more the model circumvents me by injecting horrors beyond my comprehension.
>>101848129Yeah this is one image with one prompt
>>101848129ヤにらイうこたしっこテーリ...ラド...
>>101848129
>>101848157AlignYourSteps seems kinda broken with flux, but it's fantastic for sdxl
>>101847080>Local Diffusion Generalcan i post pics prompted on paperspace lol
Loras are flowing in now. Most of them are garbage
I'm looking to buy a used 3090, what price is good these days?
Couldn't get this one to look like Miku despite putting her name in like a dozen times, but I still like it
okay .. flux LORAS starting to get good
>>101847935People who are staying on 1.5 are poorfags, there's even less chance they'd switch to flux lol.
>>101848237>Couldn't get this one to look like Miku How is that possible? If Flux is the best at anything, it's drawing Miku kek
>>101847957However LoRAs are coming out surprisingly well, it might be possible to undo the sloppiness after all.
>>101848260I wholeheartedly agree
>>101848221>>101848241hooo boy we're gettin somewherehttps://civitai.com/models/638052/ps1-ps2-old-3d-game-style?modelVersionId=713553already leagues above 1.5 and XL/Pony's equivalent LORA. Could use improvement, but wow.
>>101848241what is the recommended lora workflow cause there are a billion different options
>>101848279>Flux.1 SIt's worthless
>>101848289Doesn't the loras on Schnell also work on Flux?
>>101848289shut up faggot you're worthless.
>>101848295no
>>101848298What about Dev x Schnell schizomixes?
>>101848260https://files.catbox.moe/giyxv9.pnghttps://files.catbox.moe/yvfn90.pngBeats me, It's a massive prompt but so is pic rel and it worked fine here
>>101848146I am actually using it, and I've tried both new VAEs, spent half a day yesterday figuring it out and realized that I might be extremely retarded while taking a screenshot for you.I just didnt notice I left CFG after fucking around with SDXLTurbo. God damn it, I didnt notice anything wrong since the images produced were so grounded in reality, they were exactly what I was looking for if not for the washed out part
>>101848295ya the work on both, but I alot of testing should be done on schnell, cause I think it will ignore alot cause it converges so fast>>101848304thats just stupid attention grabing .. they saw its possible to merge em, but the results are just mediocre
>>101848165Negative prompting is what you need. >>101843586
>>101848311yeah that will do it, default is like 7. but that checkpoint is very very good for anime I have found.https://github.com/DominikDoom/a1111-sd-webui-tagcompletethis extension lets you add booru tags for anything fast and the model will yield good stuff for most booru tags for anime characters or whatever, it's really useful even if not using a lora.
>>101848283I second this>>101848279>>101848241is it feasible to run/train them on 16gb?
https://civitai.com/models/640459/impressionist-landscape-lora-for-flux?modelVersionId=716306First art style lora. Impressionist landscapes.I kind of want to train one of Greg Rutkowski just so he has an actual seizure irl.
where's this fromhttps://huggingface.co/datasets/la-ji/sd-prompt-image-in-the-wild-counterfeit
>>101848374>is it feasible to run/train them on 16gb?not that I know so.. I mean you can, but it will take ages, even a 4090 can only do it on fp8 .. if you want to do a fp16 lora you need 40GB VRAM https://twitter.com/ostrisai/status/1819802556261863925
>>101848279Definitely more ps2 than ps1. The low res warped PS1 textures is something that Flux simply cannot do, so I really can't wait to see a proper PS1 and N64 LoRA
>>101848451Would be more fitting to describe those renders as prerenders rather than ingame shots. phong/gouraud shaded.
>>101848451>Definitely more ps2 than ps1. The low res warped PS1 textures is something that Flux simply cannot do, so I really can't wait to see a proper PS1 and N64 LoRAthis, I'm also waiting this one hard
>>101848461Nice SVD
>>101848306>It's a massive promptUsing up all the tokens produces significantly better results in Flux I have found. The more schizo your prompt the better.
what causes this again?
>>101848476>low res warped PS1 texturesyou mean unfiltered.. but yea FLUX refuses to go back in time to the 90s.. the 80s it can do.. but 90s computer graphics is not there
>>101848399Do fp16 loras work on fp8? Logically I think yes, but I genuinely don't know.
>>101848399oh damn, was really impressed with FLUX but it takes so long on my poorfag 4080s (and you cant preview the output while it calculates steps, unlike SD). Well I guess no LORAs for me for the next 10 years>>101848344thanks, I'll give it another shot! used to use it on 1.5 but more often than not I had to change the words back
>>101848497It's what happen when skinwalkers forget to moisturize.
Found a node that works at high CFG without requiring DynamicThreshold: https://github.com/comfyanonymous/ComfyUI_experimentshttps://imgsli.com/Mjg2NzUw
>>101848518Interesting, can you please try a few more examples in different styles?
>>101848508>god damn 4chan glitched out and didnt let me finish my messageI wanted to ask if anyone managed to replicate picrel on 1.5/XL/Pony/FLUX? I've been trying since winter and failed miserably
>>101848533>Interesting, can you please try a few more examples in different styles?give me a prompt I'll see what I can do with it
https://civitai.com/models/640459/impressionist-landscape-lora-for-flux?modelVersionId=716306Wow this is a very impressive LoRA, added a couple of artists too.
>>101848539DALL-E gen, too advanced for Flux
>>101848518you bitch >>101848531Just do what I do and call yourself a retard and attach the pic in a reply.
>>101848320cool image i like the pyramid
>>101848503they do.. its just less numbers.. every weight is a floating point number, if you have 16bit you got 16.7 million.. if you got 8bit you got 256 numbers.. it will work.. and yea pic related was done in fp8>>101848508I actually ran OOM earlier when doing FLUX with a lora.. on my 4090 it works fine in either fp16 model or text encoder.. but If I load a lora it wants just that little bit more than 23.4GB and pushed to OOM if I ran it on FP16 model
>>101848556who dat?
there's something terribly wrong with forge's image qualityimages are nowhere as crisp as with A1111 or comfythoughts?
>>101848544I will give you some catboxes I did with DT.https://files.catbox.moe/g9getp.pnghttps://files.catbox.moe/s98qof.pnghttps://files.catbox.moe/0cpx1d.pnghttps://files.catbox.moe/0fl91b.png
>>101848611>that last gen is miku finding out about the mikuposters on the /g/ ai generals
>>101848554yeah I know. Yet my dream of local models achieving this greatness at some point lives on>>101848564I guess until chinese would release old H100 on custom PCBs with hacked drivers or maybe RTX7000, this is not the most entertaining model to play gacha at home>>101848605check samplers and settings. my gens are 1:1 on both. Though pony models RANDOMLY output colorful fog instead of images on random tag combinations, still didnt figure out what triggers it
>>101848615>They gen me doing what?
>>101848605post examples, I have had no issue with Forge
>>101848611all of em are good, but why do they have such a grainy texture?>>101848620>I guess until chinese would release old H100 on custom PCBs with hacked drivers or maybe RTX7000, this is not the most entertaining model to play gacha at homeyaa if moores law holds we got full flick making AI in 10 years>>101848560thank you
>>101848550>By Frank W Benson. An impressionist landscape of a couple fine dining at a French restaurantNot sure if following his style much but not bad
>>101848611>GuidanceNeg 10Sorry but it doesn't seem like this node can handle guidance 10, there's something special about CFG + DynamicThresholding that just makes it work somehow
>>101848620>>101848629not just with fluxshit's eternally blurry, and never quite the same
>>101848620>yeah I know. Yet my dream of local models achieving this greatness at some point lives onCook a lora from your shitting indians folder. You must have thousands by now
>>101848387i think it's from stable horde, field names seem familiar, karras=True being default, the available post_processing options
>>101848650>Cook a lora from your shitting indians folder. You must have thousands by nowsomeone do this right fucking now, make it blow up as hard as that one indian documentary that flooded the internet from /pol/ a while ago.>the next best FLUX lora could be one memeing on indianscould be big!
>>101848644Ah fair enough, will be interesting to see the prompt none the less, even without the 10neg. Particularly the Boris Valejo one>>101848636>why do they have such a grainy textureThey are over cooked with high cfg. Was early tests from anons Dynamic Thresholding workflow, settings aren't dialed in
>>101848669I have no idea how to do it though and my GPU is shit, I'll start by collecting shit rocket memes I guess. I'd love if someone release the lora for XL/Pony
>>101848655funny fact about FLUX.. DDIM does not work on CFG == 1 .. but if use the hacks .. it works perfectly fine.. all pictures like this and above related, I made em on DDIM
>>101848676>Ah fair enough, will be interesting to see the prompt none the less, even without the 10neg. Particularly the Boris Valejo oneOk then I'll get back to you once it's done
>>101848682>2nd panelFinally, after all these years I have found it... the clitorus
>>101848676>They are over cooked with high cfg. Was early tests from anons Dynamic Thresholding workflow, settings aren't dialed inyou got em like this? I am running CFG == 8 no problem
>>101848706>Was early tests from anons Dynamic Thresholding workflow, settings aren't dialed inI'm that anon, so you think that going changing "interpolate_phi" from 0.7 to 0.87 makes it better?
>>101848706Neg CFG was set to 10
>>101848725pure superstition, I like to move abit of the known settings.. but it wont make much a of a differernce, its about color depth and the like>>101848727kek, pic related
>>101848641Nvm seems overcooked
>>101848743now swap positive and negative prompts
>>101848676>stylized abstract modern artwork featuring bright colours, complex geometric shapes, highly detailed, intricate visible brush strokes, of Hatsune Mikuhttps://imgsli.com/Mjg2NzUxyou said the "Boris Valejo" one but I didn't find any workflow that has his name on the prompt
>Upon final inspection of the FLUX checkpoints, they discovered trace amounts of E.COLIS-Sars?!
>>101848743if you are writing the same text for the both of them, why won't you just fuse them, would be less a pain in the ass lawl
Oh yeah almost forgot to mention, i fixed my problem earlier with forge;Now loading an old forge for anything not Flux, and keeping the latest forge as Flux develops to the point i can actually feasibly use it.>now someone please share some funny indian gens i dont have mine even named so i cant find em
Testing https://civitai.com/models/639820/kazuma-sketchbook-flux?modelVersionId=715659
>>101848775cause I sometime like clip more than txxl .. depends on the prompt>>101848760here you go.. pic related
they should have trained it with T5 only, and proper unfiltered captions
>>101848764>you said the "Boris Valejo" one but I didn't find any workflow that has his name on the promptIt's the demon one. I must have asked an LLM to describe a BV painting and slapped miku into the prompt. Was a few days ago and I am quite retarded
>>101848832>cause I sometime like clip more than txxl .. I wonder if there's a node that does that for us already, jut pulls out key words for clip
>>101848809
>>101848841if you want any type Miku just follow this workflow, replace artist with the one you want and replace cubism with the style the artist paints in>https://files.catbox.moe/cbym12.png
Miku wearing a business suit and using two pistols, in the style of john wick:
>>101848869 havent seen one yet.. but as Black Forest anon said a few days ago when he visited us: "There was a reason why we kept clip"
>>101848876
>>101848692interesting. the ddim_steps field name was just an artifact from past development, they supported other samplers
>>101848899DALL-E 3 didn't and it beats Flux
>>101848913
>>101848636>all of em are good, but why do they have such a grainy texture?that's because of the DynamicThreshold, it fixes CFG but also gives this white/grainy effect, there's supposedly a better alternative than that but it's something that is "very slow" based on their authorshttps://github.com/scraed/CharacteristicGuidanceWebUI
>>101848939Overall not a bad LoRA
>>101848876>>101848913>>101848939>>101848953soul
>>101848841Here's the comparaison with Borus Valejohttps://imgsli.com/Mjg2NzYz
Okay, flux paired with gemma2b is better at iyashikei slice of life 4koma than most human authors.
>>101848899>There was a reason why we kept clipprobably for training purposes so people can use their existing data setsIf Flux becomes the default they get more funding, more recognition, more money. They are out the gate really fucking strong, and we have 2 years of autism knocking out loras really quickly now.
>>101848876since I can't read japanese to me this just looks like someones real drawing.
>>101848841Take back what I said, it seems like TonemapNoiseWithRescaleCFG can handle CFG + Guidance 10https://imgsli.com/Mjg2Nzgz
>>101848970Thanks, still doesn't beat DT, looked a bit burned.
>>101849048>looked a bit burned.the parameters can be changed, I used the default one, DynamicThreshold looked like shit when I used the default one, took me days to figure out the sweet spot kek
>>101849076>the parameters can be changed, I used the default one, DynamicThreshold looked like shit when I used the default one, took me days to figure out the sweet spot kekcan't really tell from your poogens
>>101849057This watchmen adaptation was fucking weird.
has upscaling always been as slow as genning? My gens could be taking half the time if upscaling didn't take forever, where i sit at 1.35s/it gens, upscaling takes 6.83 as high as 9s.it..
>>101849200upscaling has always been slower, bigger image means more work to do in each iteration
>>101849244whats this i see about model switching at different phases? Should i try a model meant for upscaling to get a bit of a boost?never really saw anyone talk about this in SDG so im pretty ignorant.
This one may be a bit too spicy for /g/: https://files.catbox.moe/ehcgqs.png
>>101849303haha just kidding haha
So are we back with LoRAs for Flux? Something change?
>>101849441Heh
the humble peugeot 206
>>101849642the not-so-humble 106 GTI
>>101849754not enough rice
>>101849765Gas gas gas!
>>101849765naisu
>>101848486Yes. Because (((CLIP))) goes up to 77 and t5 goes up to like up 500.
I love flux manga.
Is Flux any good for photographic-like images? The online version seems meh.
>>101850118There was only one, now there's two. Just like Agent Smith.
qrd on flux prompting? is it just boomer prompting or is there something else going on?
>>101848306Prompt for pic rel? Looks great.
>>101850218same as sd3
cyberrealistic pony is surprisingly good, was having issues getting pony realism to do unique poses and more than just 1girl portrait, though having issues getting consistent quality. ez to get passable and consistent results though.>i wonder if her broach getting duped sometimes is a result of bad tagging?
>>101850218It follows the prompt well unless the photo involves an adult human female. If you don't feel like writing a boomer prompt, just get ChatGPT to write it for you.
Huh, I thought there was a fix for dynamic thresholding producing white noise when used with adaptive guidance with last steps at cfg 1. I even reinstalled it (and updated everything else) and triple-checked to use the same settings that were posted here many times but alas
>>101850415does your prompt contain a stray ":" by any chance? Also, every recent model inserts eyes or proto eyes in places where they don't belong. I'm not sure why.
post you're gen when /ldg/ won
Will I ever be able to train Hunyuan Loras on a 12gb?
>>101850500lol no
>>101850480No but i just noticed my current gen accidentally left a comma in lmao oops
>>101850415One sign of a braindead model (most fine-tunes are like that, sure sign is when its author makes a new version every two months...) is the inability to produce anything else than instagram style human portraits.One way to gauge a model is to test common painters.
>>101850177I don't get anything even close to that.
>>101850558Yeaaahh I just wanted to give that model a break because it's the best of its kind, that one amateur realism lora is shockingly BAD. looks like a deepfried amalgamation nightmare and it's highly rated.Cyberrealistic was always hit or miss, hit it was of one of the best models for 1.5 and a kinda bad XL model. In this case for Pony i still can't judge because i'm using it for the first time.It's certainly higher quality than any of the other cyber realistics.
>>101850526:( vramlet life sucks
>>101850627What model are you using, that's not Flux.
>>101850629Seems alright.
>>101850675No. That's SD1.5 Humu as in the filename. This is what I get on flux shnell
Hey guys i'm in need of help with a inquiry in how some Ai videos are made First, are you familiar with "dreamcore/analog horror/liminal aesthetic? Well if you are not here are a few exampleshttps://www.youtube.com/watch?v=2vtn9EjSlJ0This is DeadTempo visions he does dreamcore stuff mostly https://www.youtube.com/watch?v=G915hZm18qMGlumlot, same dreamcore aestheticshttps://www.tiktok.com/@fearstfilm/video/7368984349745679658?q=fearstfilm&t=1723469007140FearstFilm, he does analog horror liminal spaces stuffso, my question is how are those videos made? i mean what AI app or software is being used...i was thiking at first MidJourney picture that then is animated using a video ai software like KLING but i have been trying to achieve the right aesthetic and i just can't so there must be something i'm missing. since DeadTempo and Glumlot achieved almost the same aesthetic i guess the procedure is something entirely diffetentany ideas? much appreciated (and if not well maybe direct me in the right direction where i could find some answers and i'll be on my way)
>>101850787use the realism lora, or find the right keywords
>>101850629In my experience pony is horrible, it does characters well but even then it needs wrangling. Have never tried any photographic pony fine-tunes (in my logic that's useless because pony is a cartoon model anyway but what do I know).
>>101850787Humu is actually very good with artists and art styles, better than some other models. It's something you wouldn't believe.
>>101850799or this oneI'm using the same prompt. It's not about the prompt, it's about the image quality. It's not the same as >>101850177So wondering what's the trick. I'm installing locally. Maybe the only version is nerfed.
>>101847080Can someone please try generating the following with FLUX:"backgammon board, view from above"or just"backgammon board"none of the other image generators, even Midjourney never got this right.Wondering if they have progressed anywhere on details with FLUX
>>101850820welcome back
>>101850853>It's not about the prompt, it's about the image qualityyeah, use the realism lora
>>101850853*on-line
>>101850446And this >>101849076 thing gives me the same shit. Guess it's the problem with adaptive guidance then, even though I have the latest updated version. Does it need any fixes or special settings applied? Using adaptive guidance with thresholding of 0.994, cfg of 6 and uncond thingy at 0. Tried tinkering with valus but nothing changes Removing adaptive guidance altogether gives me overcooked shit
It's already heating up and that's before I even opened up the oven door to find some piping hot fresh bread...>>101850883>>101850883>>101850883
>>101850830pony blows every other model out of the water. It requires wrangling but you know that going into it. You can do things with pony that are outright impossible with flux. Flux is good for writing and memes.
>>101850925I think you should reinstall AdaptiveGuidance and DynamicThreshold, because it's working for me
>>101850861thanks, but I am not a regular
>>101850925>And this >>101849076(You) thing gives me the same shit.that's normal for that one, that TonemapNoise shit needs to be deactivated when the CFG returns back to 1, a bit like DynamicThreshold did a few days ago