Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101861781>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
>>101867704>picked the miku in a bottle I genned on the train at 512x512 on my laptop but not my 500 step brapper
>>101867729>loras trained with small datasets means it already knows the concept!!loras have always been trained with small datasets lol
The cock LoRA got real good real fast
>>101867791To the anon who shared this catbox, this is a good fucking prompt. Boomer as fuck but you can't argue with the results.
>>101867791this, seeing anons talk about Flux loras being so awesome with "ONLY 20 IMAGES!!!" is funny
Blessed thread of frenship
>>101867862rompt?
>>101867680>They were so excited by the VLM that they overlooked this.they are literally authors of latent diffusion and vqgan, assuming stupidity on their part is unwise, especially seeing how they went an extra mile to wipe some artists out of the dataset, and they're from EU with their lawsthey definitely know what they're doing and it's definitely on purpose, I'm not sure why though
are flux loras fucked or something? works perfectly on 1:1 ratio but i switch to 16:9 and its all fucked upusing this https://civitai.com/models/639820/kazuma-sketchbook-flux?modelVersionId=715659these are both on 3.5 guidance and lowering it for the 16:9 gen helps but loses a lot of quality
> in the style of famous art that bears the "Rembrandt" signature
>>101867923quak
>>101867918Because traditional art is white culture, unironically. They want to genocide white people. Future generations won't know there were white artists. When a rebel try to teach the kids about Picasso and van Gogh, the kids will tell them they must be making it up, because they asked GTP 76.6 about them and then tried to prompt them on Flux 46.1 and nothing. They will tell the future generations that blacks invented everything. Get ready.
>>101867894nvidia brand robot having an existential crisis whilst doing laundry
>>101867918Because they don't want to get sued or run afoul of the EU AI act. It's probably the same reason it sucks at female nudity.
>>101867923it's not specific to flux. any lora trained on a shit dataset will make "fucked up" images at resolutions that were not present
>>101867933Same prompt but with van Gogh.
>>101867955please.>>101867923it is all a bit experimental
>>101868168Miku would never
has anyone been successful in getting a woman to suck a penis with the penis lora? I know, crazy idea, call me fucking van gogh
>>101868187Anon we have barely just attached nipples and cocks.
>>101868187lol
>>101868219now attach mouth to cock lol
Hey guys post strawberries, I will award the most creative one with a reddit gold
>>101868174For you
>>101868236ok well now i hope we dont get it
>>101867923historically it's been the case that if there's no variety in the lora dataset then there's limited variety in the final outputs
>>101868236pony-like miniature adult body. doesn't look anything like a child... it's the natural outcome of trying to "sexualize" them when the training data does not have overlap between children and sexy. it doesn't know what to retain from both concepts when it merges them together.
>>101868493I can take it. brazilian babyfat bubble butts, dude.
>>101868605>waaaaah be my divorce rape slave!No. Learn to compete like men have had to do for the past 1,000,000 years or enjoy your cats.
https://www.youtube.com/live/fxA6svitx5Y?t=5158s>"in one sense this hearing has been very interesting - you are all advising us to do nothing"lmaoAny antis waiting for the government to step in are delusional. There's three 50 year old karens there as experts, exactly the type of person to be anti-ai, and they're still not asking for regulation.
>>101868605You are not special, Karen>0 sexYou realize escorts are a thing, right?
>>101868767>You realize escorts are a thing, right?>has to pay for sexlmfao, you really showed him
>>101868807If you are a normie, you always have to pay for sex one way or another, dinner, Motels and fuel are not free
>>101868850is that the cope you tell yourself?>motels....lmfao
>>101868862>is that the cope you tell yourself?Yeah bro, I guess you are right. Women, even attractive ones, magically spread their legs for you, for free, without you ever needing to do anything for them, and they take you to their place to fuck. And you don't even need to be a chad, it's automatic. Noted.
>>101868850Lmao loser detected
how do I get me some bitches for free?
>>101868957Be at least somewhat nice to have aroundGo outsideMeet new people from time to time
>>101868957Just be yourself and just show up in public bro, you don't even need to do anything or pay anything, always work, as seen with the slayers ITT
>>101869016I'm sorry you're ugly saar
>>101869027I had landwhales hitting me up IRL before, but I still prefer my hand over them, tks
is the multiGPU + Loras bug fixed?
>>101868584kek, incels fucking wish. Women have you all by the balls. You’ll come home one day to find your robot gf eating out the bitch who laughed you off.
Imagen-3Not bad
>>101869147but didn't you see that the incel depicted himself as handsome and thin and happy? clearly he's the perfect man and its the women who are wrong for not wanting him
I just find a bit ironic there are people bragging about sex life in a thread where loners post AI 1girl waifus, of all places, really makes you think
>>101869195People can have girlfriends and watch porn too, your point?
>>101869164YES>>101868942very nice
>>101869210See, if you even have to jack off while having a girlfriend available to offer you pussy, means there is something wrongI doubt people who date a hot stacy would rather jack off to AI gens
stfu Teebs
posting gens, hope nobody minds
>>101869295>>101869257UUUOOOHHHHHHHH
>>101869222thanks
>he's gaining
>>101869164nf4 Flux.dev
wat
>>101869694weird
>>101869946>Mainland, Taiwan and artificial land on SCS.
>>101869706>>101868282asking my questions here.why did the general split? it's a hassle
>>101868937>magically spread their legs for youit's not magic, it's mutual attraction. I understand the concept must be extremely foreign to you if no one has ever found you desirable.
>>101870159probably something fishy about the custom node
>>101870159evolving beyond a single company, other thread has served its purpose but is now a relic
>>101870159sdg hasn't been relevant in months >attaches avatar
>>101870159referring to all image generators as "stable diffusion" is as dumb as normies referring to all text AI as ChatGPT
>>101870184but i found that i can't even restart comfy through the manager due to this...
>>101870159two schizo factions
>>101870255you're probably doing something wrongthis guy will show you howhttps://www.youtube.com/watch?v=L5BjuPgF1Ds
>>101870169then don't act like it's the guy's fault to have to pay for sex or have to go through several hoops to get pussy (like 80% of the male population do), not everyone is a chad and not everyone is into ugly women
>>101870159The thread split because some people did not want to have to post in a thread with a certain individual.
does Flux work with TensorRT btw?
>>101870141>>101870093>>101869927>>101869902fuck you, now I have to stay and collect all the monke pics
>>101870751anyone who actually knows though? someone who tested it?
>>101870817based
Anyone know why I'm getting these weird seams (look at the top right, towards the wings) when latent upscaling and resampling? Doesn't appear after the upscale, only after the resample.
>>101870535>then don't act like it's the guy's fault to have to pay for sex or have to go through several hoops to get pussy (like 80% of the male population do)it is the guy's fault, if they weren't such simps, women wouldn't act like princesses in the first place. If you treat her like a celebrity, she'll treat you like a random fan, simple as that
>>101868850the other guy is LARPing levels 9001929, but I'll tell you that I had a woman pay for everything for a few years, basically cause I was an uni student with no job and she was older than me and employed. so no, don't take the jewish propaganda posted on 4chins as 1:1 description of reality, it's a trap
>>101870995define "upscaling and resampling". ultimate sd upscale? raise the mask blur, try 32 and no seam fix>>101870817this is a very nice gen, gives me something
>>101871102>I'll tell you that I had a woman pay for everything for a few years, basically cause I was an uni student with no job and she was older than me and employed.damn you must be really good looking anon, no homo though kek
>>101871158it was his mom
>>101871102Then you are not a normie, I said "If you are a normie">>101871023Unfortunately even average-looking women have a diva complex these days, so non-chads have to go through some hoops to get pussy
>>101871214but she was paying for it...
>>101871214>Unfortunately even average-looking women have a diva complex these daysthat's because men treat them like princesses aswell, average women gets hundreds likes on tinder in a single day, no wonder they think highly of themselves, men just can't help themselves but to simp for every female that exist, even fat ugly bitches, they have absolutely zero standards and we live in a world where morbide obese women wins more money than a fucking surgeon :)https://www.youtube.com/watch?v=H-SM85I3UvQ
>>101871275how much does the average OF thot make in a month, anon.
>>101871302>the average OF thot make in a monthpeanuts, it gets so bad that they're offering their pussy irl through it to pay rent, off course x, y and z are mad rich, the rest, not so much
>>101871302How much matches an average thot make in a single day on tinder anon? https://youtu.be/lvSX3DZr_FQ?t=21
>>101871358ESL anon why did you change my question?
>>101871376>can't even understand the topic that was discussedit wasn't about a prostitute's salary, it was about the insane amount of attention ugly bitches have, poor reading comprehension ESL anon >>101868850
being an absolute retard slopping it up training a Yoshiaki Kawajiri flux lora w/ sd-scriptsany research on how much better llm classified captions are for loras yet? just using my existing tagged dataset for now
>>101871421>>101871275>we live in a world where morbide obese women wins more money than a fucking surgeon are you retarded, retard anon? you might be slightly mentally deficient, retard anon.
>>101871452>seething this hard because his ego's too fragile to admit defeatlike clockwork
>>101871480bait
Here's hoping all you incel anons grow out of your shell and meet the one who will simp back for you. Also post gens, is this irc
>>101871508>Here's hoping all you incel anons grow out of your shell and meet the one who will simp back for you.love and relationship is a meme, it's just based on superficial criterias like being young, beautiful, rich...
>>101871508I am in relationship with my waifu
>>101871508bearlets when will they learn
For those having OOMs during lora loads after recently updating ComfyUi, have you tried this PR? Did it fix anything? https://github.com/comfyanonymous/ComfyUI/pull/4319
https://civitai.com/models/643197/dark-helmet-flux-d?modelVersionId=719450Kino inbound
is flux like pony
Rey Flux lora when?
>>101872014>Flux-DI'm glad people are focusing on Dev, that's the good version to improve on
>FP16
>>101872014>still doesn't work on nf4Man, poorfags got filtered hard
>>101872184>NF4
>>101872193>Man, poorfags got filtered hardeven richfags are getting filtered at the moment, fucking OOM on lora loadings because ComfyUi fucked something up on his latest commits >>101866190
>>101872184>>101872221I can't even tell which one looks better, only that they look different.
>>101872184>>101872221>hard shadows on nf4I realized that the less quality a quant have, the harder the shadows have, a bit like video games, to get soft shadows you need more accurate lighting algorithm
>>101872193Poorfags got an option to run flux at decent speeds instead of waiting for 5+ minutes, still a colossal win t b h
>>101872291also worse subpixel detail
What UIs can use Layer Diffusion? Only Forge and Comfy?
anyone use fooocus? im trying to img2img logos and this guy seems to have a way to do it. i have not had much luck with sdxl, it screws it up too much https://www.youtube.com/watch?v=MnGYEB_oOqE
https://reddit.com/r/StableDiffusion/comments/1erbumy/chinese_are_selling_48_gb_rtx_4090_meanwhile/is there a way to buy those cards through internet though? would be cool
>he doesn't eat spaghetti with his hands
>>101872714They might at some point filter their way to the West because of the VRAM amount but realistically at the prices China is buying them for, the reimport price is better off being spent on actual Enterprise cards.
>>101872539Best inpainting option out there is the krita ai plugin. Uses comfyui as a backend, supports controlnet and other cool shit. Layers, regional prompting, transparency support, you name it
>>101872853hamdurudubidubidurula, I love eating spaghett with my hands right after I wipe my asshole with it
>>101872970Oh shit, I didn't even realize there was a flux support update, shame it's not nf4 though
>Runge-Kutta samplerhttps://github.com/wootwootwootwoot/ComfyUI-RK-Sampler?tab=readme-ov-file>Restart samplerhttps://github.com/ssitu/ComfyUI_restart_samplingDid someone try those samplers on flux? I don't think Euler is the optimal sampler to use (I don't want slow shit like Heun though)
>>101873108Heun, just like DPM++ SDE and some others, needs half as much steps to cook compared to regular samplers, so it takes exactly the same amount of time
>>101873108give deis a try with beta or ddim scheduler
>UI_model-name.png
Flux seems cool I'll come back in a month when people figured out how to make gooning material
/sdg/ self-destructing
>>101873143>Heun, just like DPM++ SDE and some others, needs half as much steps to cook compared to regular samplers, so it takes exactly the same amount of timeBut Heun 20 steps gives some similar pictures than Euler 40 steps imo so I don't see the point>>101873152thanks anon I'll keep that in mind
>>101873108>Runge-Kutta samplerExcuse me?
>>101873108https://stable-diffusion-art.com/align-your-steps/>Diffusion models generate samples from noise. This process is called Reverse Diffusion, and we need to solve a reverse diffusion equation. Since we cannot solve the equation exactly on a computer because it has no analytical solution, we break it into discrete steps to solve it.Damn, imagine one day some math nerd finally resolves the equation, that means no more samplers shit to choose, and it would be really fast to make pictures
>>101873182I don't think Flux is gonna be better at coomer shit than pony anytime this year. Maybe by summer of next year. Assuming no better base model releases before then that would steal all the momentum and interest.
any tricks for speeding up t5 encoding on older Ryzen cpus?
>>101867704Does flux recognize Akko yet?
Do you need 24gb of vram for fp16? What's the best option for 16gb vram, nf4? fp8?
>Prompting "Indian man" gets men that are mix of pajeet and native americanDamn can't believe Flux dataset is racist and Flux is getting cancelled.
>>101873846>Do you need 24gb of vram for fp16?yes>nf4? fp8?yes, most will prefer nf4 due to size
>>101873152>give deis a try with beta or ddim schedulerYou're right anon, deis is pretty good
> Looking out from the cover of a manga that takes the appearance of the distinctive artist style of "TORIYAMA" is the likeness of a young woman.I like how t5 lets you just write absurdly long statements like this. Not sure this is quite TORIYAMA style though. Good example of how important prepositions are to prompt comprehension.
>>101874006>I like how t5 lets you just write absurdly long statements like this.is t5 still the SOTA of text encoders? I was suprised they used it for flux even though that model is 2 years old
>>101874047>was suprised they used it for flux even though that model is 2 years oldwell flux wasn't trained overnight. It's not like they could have used something a month old.
>>101874063>well flux wasn't trained overnight.I doubt it took 2 years though, but that's a good news for the image gen community, we managed to get something close to dalle in terms of prompt understanding with a 2 years old text encoder, imagine if we used something more shiny, gemma2-9b would be a good candidate, this shit is smaller than t5 (11b) and just better overall
>>101874000omegachecked, happy to be of help. i always end up going back to deis myself
>>101874047hard to improve perfection
>>101867704woah 2 of my pics have made it into the OP.feels good man
>>101874130deis + ays looks like a good combo too
t5 fp8 takes 80 seconds on my machine. that does, admittedly, include loading.I need it to take less time. Can I force ComfyUI to keep it loaded in RAM? What else can I do?
>>101867923The author didn't mention the trigger word, did you try "By Oda Kazuma" at the beginning? Took me a while to figure it out, though I haven't tried other resolutions
>>101874176I like ays+ more
>>101874201>Can I force ComfyUI to keep it loaded in RAM? What else can I do?yes you canhttps://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101874241>I like ays+ moreI managed to find some documentation about ays, but not about ays+, what does the "+" mean here?https://stable-diffusion-art.com/align-your-steps/
>>101874221h..h-hot
>>101874267>reading documentsGet out with that nerd shit, I just click stuff and compare the results
>>101874312>Get out with that nerd shit, I just click stuff and compare the resultsthat's a good way to pass some time while waiting for a 1mn gen to be over kek
What the FUCK is wrong with lora makers not describing how to use their loras? I was looking up facial expression loras for SDXL. Seems like this one might be good (https://civitai.com/models/541620?modelVersionId=603774). Except how do you even use it? Model card has no info. Example images are schizo ESL prompting that doesn't even stick to any one consistent style of triggering the lora.Like if you're training a facial expression lora, surely you would just caption each image normally, then add "crying", "sad", "happy", "excited" etc as tags to capture the expression, right? Simple, obvious, don't need any global lora trigger phrase, can't go wrong. But nobody fucking does that apparently.And before you criticize me and say I don't know what I'm talking about, I have trained complex concept loras and they are nowhere near as hard or confusing to use as this shit that I see all the time.
>>101874334what do you expect from people who train loras on 1070s in 2024?
>>101874334>What the FUCK is wrong with lora makers not describing how to use their loras?Playing Devil's advocate, as an engineer, making documentation is easily the worst part of my job, I much prefer experimenting rather than writing some detailled tutorials for the retards, but I'm doing it anyway because I know the users didn't spent weeks on that shit so they aren't much experienced and need some hand holding, still hate doing that shit though
That guy nailed it with that Emma Watson lora, the other real people loras currently out for flux don't come close
remove rei and asuka from every ai model so weebs have to be more creative
>>101874334Can't you look at the example picture tags and figure out?
>>101874334Why even upload loras in the first place
>>101874384you know I am surprised some autism fight has not been started over a major model maker removing just one of them, I suppose the ponyfucker removing vtubers was close
>>101874406>Why even upload loras in the first placethis, if Flux was pretrained proprely, we wouldn't need loras because it would've known most of all the concepts
>>101873756akko specifically got removed from the model
>>101874375Does it contain the training parameters?
>>101874397Ok let's take a look.>cinematic film still of <lora:add-detail-xl:0.3>0 <lora:detailed:0.3> realism <lora:perfection style:0.2> perfect <lora:eyetune:0.5> <lora:xl_more_art-full_v1:0.1> <lora:Ron Fricke style:0.2> <lora:Unreal Engine style:0.3> Unreal Engine style <lora:TeethXL:0.3> <lora:Chiaroscuro Lighting Style:0.3> Chiaroscuro Lighting Style <lora:facial expression style v2:0.9> <lora:crying style:0.9> crying style teary crying facial expression of a woman with a red scarf is crying with huge tear drops, movie themed, sharp, detailed skin, epic cinematic photography, artistic, creative, wrinkly face, dramatic lighting style, cinematic color style, Kodak film style, skin pore, emotional style, gustier style, facial expression style, teary crying facial style, solo, looking at viewer, smile, short hair, brown hair, 1boy, closed mouth, male focus, cum, facial hair, portrait, meme, realistic, black hair, blurry, parody, asian, photorealistic, teary eyes , shallow depth of field, vignette, highly detailed, high budget, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainyoh, yes, very obvious how to use itwhy not just use basic tags like i mentionedall these people are incompetent lora trainers
>>101874454Redditspacing
>>101874477Exhibit 101 of:>when the sage points at the moon the fool looks at the finger
>>101874477>formatting posts for readability or separating lines for emphasis is le reddit behaviorkill yourself
>>101874477>newfag who can't read trying to be cool
>>101874492>>101874511>>101874525yourconcessionsaccepted
>>101874454>why not just use basic tags like i mentionedIkr, why can't it be as simple as writing "Emma Watson" when using a Emma Watson lora for example, why going for convoluted trigger words, that's not natural at all
>>101874528>concessionsyou mean that one? >>101874477
>>101874176how do you have 'ays' available in the basic schedulers node?
trying out the flux realism lora
>>101874548I don't know man, at this point I downloaded so many custom nodes, it just appeared at some point...
>>101874548It should already be there?https://github.com/comfyanonymous/ComfyUI/commit/644a3ae58d426ffbbc02ef4104034c98e8fc6513
>>101874530retards think using the real name will taint the learningit don't, unless "Emma Watson" generates a chair or something in the base model
>>101874548I got you, it's this shit: https://github.com/pamparamm/ComfyUI-ppm
>>101874573damn, tell me if you find where it came from. I'm surprised AYS even works with Flux effectively, since the native node on Comfy is a separate node (>>101874594) and doesn't work currently..
>>101874608>retards think using the real name will taint the learningit's the opposite imo, Flux already knows the concept a little, adding the lora will provide a boost, it's retarded to reinvent the wheel
>>101874617ty anon
>>101874639yes, T5 knows what female name looks like and it's a great starting point they throw away to start from a vector it has never seen
>>101874662>yes, T5 knows what female name looks likeit even knows who Emma Watson is, abeit not that much, but you can see it has seen some pictures of her at some point
>>101874684>>101874693>>101874702>>101874715cool
>>101874623>I'm surprised AYS even works with Flux effectivelyeven Euler CFG++ works on flux, even though it's worse than Euler imohttps://reddit.com/r/StableDiffusion/comments/1eq98ca/euler_cfg_actually_works_on_flux/Flux has way more potential than what we initially believed, we thought lora training was impossible because it's a distilled model and now here we are, almost all the SD features work on flux aswell
>>101874725I know, right. Flux is really, really good at creating aesthetically pleasing pixel art.
which scheduler are you supposed to use with DEIS?
>>101874793>which scheduler are you supposed to use with DEIS?>>101873152>>101874176
>>101874241>I like ays+ morewhat about the "ays30" and "ays30+" ones?
>>101872685heh good one
Recognizes Super Nintendo but not Genesis, SEGAbros...
Deis has the tendancy of overburning the black people kek
Comfy is now not clearing RAM after moving the unet to VRAM. fuck's sake
>>101875000Flux can do what SEGon't
Why the fuck are there two Stable Diffusion threads on this board
>>101875038do you see many Stable Diffusion gens in this thread?
>>101873152>>101874176>found this node linked from the other repo to merge sigmashttps://github.com/Extraltodeus/sigmas_tools_and_the_golden_scheduleroh yeah, it's schizo time
>>101875038/ldg/ is for all local diffusion models, not just Stable Diffusion
>>101875054kek it's going a bit too far for my taste, I hope you'll find some good combos though
>>101875000Here's your Genesis bro
>>101875053It is all the same bullshit
>>101875054the most i do is split sigmas
>>101875000get the mega dreve
>>101875038One is now a designated containment thread for a special individual
>>101875129what does splitting sigmas do?
>>101875129what the sigma?!
It's a shame that "ays" and "ays+" doesn't work on CFG 6 (DynamicThresholding) + GuidanceNeg 10 because on its own it's a pretty impressive schedulerHere's Deis + ays+ on CFG 1:>Hatsune Miku with dreadlocks and a light black skin skatebording, 50's comic book style, her text bubble says: "Hard to keep me in style huh?
>>101875156i mean i'm using it to simply skip a few steps for a different output.you could also use it like in this thread:https://www.reddit.com/r/comfyui/comments/1d1x5ek/dual_prompting_with_split_sigmas/
>>101875075I'm just sticking with the two sigma merges for now, but it's producing interesting results already. Despite being very different gens, they come together really well.AYS seems to be a very divergent from most other samplers.
>>101875290You compare a sampler (Euler) against a scheduler (ays+)?
>>101875308Oh my bad, it was the beta scheduler merged with AYS+. Both on Euler.
>>101875246replicated the workflow>picrelusing 1 prompt for the first half of the steps (high sigmas), and a completely different one for the 2nd half (low sigmas)
>>101872558damn, she looks like my ex
>>101875367A realistic Pony finetune?
>>101874438Rust in pupperino my sweet princess...
>>101875462She must have been a cutie
>>101875471nvm, I missed >>101874565How did you describe these clothes?
>>101868584Why do the imagined robots always look so meh.Have a perfect looking face with antenna, so much better.
Oh okay, so ays+ needs many more steps to even get close to a complete looking image.https://imgsli.com/Mjg3MzMxWhich is funny because I'm pretty sure that's the opposite of it's proposed purpose.
>>101875559Reality is way more complex than some theorical suppositions based on some "in surface" test, at the end of the day to know which sampler/scheduler is the best, the best way will always be to test them out by yourself
>>101875586>at the end of the day to know which sampler/scheduler is the best, the best way will always be to test them out by yourselfThat's exactly what I'm doing, anon...?
What does "eta" and "s_noise" means in eulerA?
>>101875606all I'm saying is you need to do your own research
>>101870587Tested it a few days ago, the workflow kept crashing.
>>101875702thanks for the info.in the meantime i found this thread on huggingface. apparently its a onnx issuehttps://huggingface.co/black-forest-labs/FLUX.1-dev/discussions/24
>>101874221Lovely hinted bra
>>101875647Once I find one that produces good results I never switch. I am still on euler for flux, can't be bothered testing the others. With it working in forge though an xy is easier so maybe I will
So what's new in 48h is basically LORAs for flux and that's it?I guess no magical multigpu support yet?
>>101875759>With it working in forge though an xy is easier so maybe I willxy plot on comfyui is ridiculously easy, when you know what custom node to use: https://github.com/CYBERLOOM-INC/ComfyUI-nodes-hnmr
>>101875812Yea but offload the clip model is good enough for me personally. If you have 2 3090s you can generate a lot large images and batch more them with the clip model offloaded.
>>101875812I have generated thousands of the same womans ass, so that's not really nothing
>>101875812>I guess no magical multigpu support yet?what about this? https://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101875850It's so much simpler in Forge for my retard brain. Like what is all that scripting shit in that box on the left? I don't know
>>101875861
>>101875882>Like what is all that scripting shit in that box on the left? I don't knowit's some shit you don't need to touch at all lol
>>101875856The car in front of her is coming head on... she's going to die.
Did they figure out how to use a lora with nf4 yet?
>>101875856prompt?
>>101875884picrel>>101875890OK but Anon I really really need you to understand just how retarded I can be. My brain is very smooth. But it's undeniable that Forge is much easier to do an XY in.
>>101875935so what was the consensus on this one?I saw it earlier and imo there isnt all that much difference between 100 steps and 500 steps.
>>101875038The last time I asked this, the general schizo'd out and tried to report me
The bakery has produced the next loaf of fresh bread...>>101875939>>101875939>>101875939
@101875961nig*bo
>>101875952Yeah pretty much. Diminishing returns begin at about 50 steps and there's little point going beyond that, but 50 is definitely better than 20.
>>101875898the prompt is based on this songhttps://www.youtube.com/watch?v=u9WgtlgGAgs>>101875920Panorama 70s exploitation movies Picture of a Black 1974 Cadillac Eldorado open-top Coupe driving in a dry desert on a highway with a woman dressed as Hatsune Miku with blue twin tail hair standing on the driver seat wearing a thong and facing away from the viewer and her arms stretched out driving on the wrong side of the Road towards a White vintage van directly in front of her on collision course, the license plate on the cadillac has the text "/ldg/"
>>101875973but nice ass tho, what are some good prompts to get a good ass gen?
>>101875851Using the whole 48GB would be nicer and allows less hassle in general.>>101875861Waiting patiently for flux to understand nsfw/risqué positions.>>101875868https://github.com/neuratech-ai/ComfyUI-MultiGPU
>>101876104>>101876234very nice work
Someone bake a new thread so this guy has somewhere to spam his mobile game screenshots
>>101875909Doesn't look like it. Since, if I understood correctly, the author for the nf4 loader custom node uses the code from forge you can check the issue (https://github.com/lllyasviel/stable-diffusion-webui-forge/issues/1001) on their github for progress updates.
