Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101709433>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
>>101715949You dropped this
official pixart bigma and hunyuan finetune waiting room, now with a participation award for lumina.
>>101715910It must depend on prompt complexity, because this was a shorter prompt
how much worse is schnell than dev
>>101715949>includes trani and cumrag>doesn't include trumpsamaop is an sdg shill. we should just rebake
>>101716119>all images are flux>op must be sdg shill because of thisBased schizo retard.
>>101716114Not much. It'll have the same limitations as turbo though.
>>101716168Absolutely haram, why is some of her face visible? Whore leaves nothing to the imagination.
>>101716157>>all images are fluxbottom left is pixart
>>101716119yeah he includes avatarfags no matter what they post
Anyone know how to make Flux do things like scanlines, monitor glow, etc? Trying for something strongly retro
>>101716385try feeding an image of what you want into something like cogvlm2, good way to learn how to prompt it
>>101716385I wrote this:>Old computer screen with a retro game on it, at night, the computer screen takes up most of the image, the screen emits a faint glow, tYes I accidently typed in the t
>>101716495Forgot image
>Old computer screen with a retro game on it, at night, the computer screen takes up most of the image, the screen emits a faint glow, game screen reads "game over" with a skeleton holding a scythe below it
>>101716450>>101716495>>101716503Thanks for the tips. Do you use the "cogvlm2-online.cogviewai.cn" website and have to do anything special? Waiting on a gen before I can try its suggestion, but picrel (made in DALL-E 3) is the style I'm trying to make in Flux, for reference.>90's retro Japanese hellish dystopian post-apocalyptic chartreuse-verdigris overtones navy-blue-gold undertones first person dungeon crawler RPG featuring (makeshift wooden corrugated metal shack) in front of brutalism concrete modern architecture in background, dark lighting and sony trinitron scanlines, screenshot of dungeon crawling game, sega dreamcast nintendo 64 crt screenshot, menu on bottom of screen with health bars, map in corner of screen
>>101716540>Do you use the "cogvlm2-online.cogviewai.cn"yes that's what i use as well. i just ask it to describe the image and paste that into the prompt, slop and all.>This image depicts a nostalgic gaming setup reminiscent of the late 1990s to early 2000s. The centerpiece is a CRT (Cathode Ray Tube) television displaying a video game, suggesting a console or a computer running a game. To the left of the TV, there's a stack of video game consoles, possibly a Nintendo 64, with a controller connected to the TV, indicating that the game is being played on one of those consoles. Below the TV, there's a VCR (Video Cassette Recorder) and a DVD player, suggesting that this setup was used for both gaming and watching movies or videos. On the right side of the TV, there's a shelf filled with video games, organized neatly, indicating a collection that the owner is proud of. Above the TV, there's a poster of the character Mario from the Super Mario series with a humorous message, and to the right of the poster, there's a framed picture of a muscular individual, which could be a celebrity or a personal photo. The overall ambiance of the room is warm and cozy, with a wooden desk and a comfortable chair, suggesting that this is a personal space where the owner spends a lot of time gaming and watching media.this prompt is probably way too excessive but making it more concise gave me worse results
>>101716658This is what it gave me>a digital artwork of a futuristic cityscape with a central building that has a makeshift appearance, surrounded by water and other architectural structures, using a color palette dominated by shades of green and blue, and incorporating a digital interface at the bottom with game-like elementsThe image/composition itself is strikingly close but the style/aesthetic is still very clean, one step at a time I suppose
I get much faster prompts switching to fp8 weight type and for the clip model, otherwise it's super slow.
>>101716686but if you have a 4090 and 64gb ram, fp16 is ideal. I'm on a 4080/32gb right now. Still, fp8 results are solid.
>>101716685you could also feed the prompt into an llm and ask it to change the prompt according to your ideas
>>101716215lolita cringe, even worse with that filename.>>101716624really nice gen
>>101716699fp16 gonna work on 24vram/32 or oom?
>>101716732I can do fp16 but it's slow cause it tries to load stuff into physical memory too plus the 32gb is basically full, fp16 is for 24gb primarily imo, but dropping it to fp8 doesn't seem to affect quality significantly (so far)
>>101716732it works on 16gb it's just slow. FP8 is completely fine and fast. It exceeds SD3 and XL by a lot.
>>101716732fp16 model with fp8 text encoder works on my 4090, both is OOM, also works vice very, fp8 model, fp16 text encoder
>>101716746>it tries to load stuff into physical memory too Wasn't there something you could do with nvidia drivers to turn that off? I recall anons having issues with lowered speed after an update until they disabled it. CUDA policy or something. This was months ago so I have no idea if things have changed since then.
What kind of vram specs do I need to run flux locally? 8GB okay?
>>101716775yes this is in the drivers by default turned on since 535 NVidia driver, you can turn it off in the NVidia control under general 3D options>>101716805if you have 64GB of RAM, an NVidia card and alot of time you dont need to use your computer you can, but it will not be much fun
First fine-tune for flux >https://civitai.com/models/621563
>>101716819sus
>>101716829>DOWNLOAD HERE :Huggingface linkBruh
I can official say Flux has replaced many types of artists. They are now worth less than a penny each.
>>101716816Okay so nobody with a normal GPU is running to locally? I haven't read up much on it, thought it was usable by more people on a standard machine. Sad.
>>101716829Click the hf link.
>>101716847i think 12gb vram is the minimum
>>101716847It runs fine on my 12GB 2060, takes about 6 minutes per 1024x1024 image though.
>>101716746>>101716760>>101716770thank you for the answers. I do have an old 2080 collecting dust here, almost tempted to go the double GPU route. I saw a post mentioning the possibility of offloading the text encoder to a 2nd device. poor PSU tho. also sorry for nogen, soon..
>>101716862fp8 weights and fp8 clip model speeds things up a lot
>>101716861>>101716862Someone ran it on a 6 GB 1660
>>101716876Any quality comparisons between the two? And is this the fp8 model you use?https://huggingface.co/Kijai/flux-fp8/tree/main
>>101716847I'm getting ~2 minutes per gen on a 3060ti with schnell in fp8. Need to test dev with fp8 next.
>>101716863>also sorry for nogenthere's no need to be, talking is encouraged in /ldg/
>>101716847many do, with 24GB VRAM you dont need to swap into system RAM and it actually runs very fast (on 3090s, 4090s) .. anything below will have to swap into system ram and it gets super slow>>101716861it works on less if you are masochistic or have a second computer you dont use, but gen times of 15-60 minutes are normal then for a 1024x1024
>>101716897I use the 23gb unet model, but I use the fp8 weight setting and fp8 clip model.
Hopefully a pruned model.comes out for flux to reduce the size and enable more to use it
Anyone know what the difference is between fp8_e4m3fn and fp8_e5m2?
>>101716819>No model details>no training info>download is off siteNot shady at all
>>101716967yea just like that wonly "merge" of flux with pony that appeared yesterday on civitai, either these ppl are malicious or seriously dumb
>>101716955If I understand it correctly (I don't) E5M2 is for more variation E4M3 is for more accuracy.
What's the t5 text encoder parameter size and version in Flux?
>>1017168985:35 on dev fp8 with 8GBAt least I got a fish.
>>101717050I'd fuck the hell out of that fish
>>101717001>E4M3>E5M2I'm a fan of E2M6 myself.
For some reason Flux has a tendency to not recognize borders/separations in a way that I don't even remember 1.5 doing, like this maid's headdress exiting the monitor
>>101717083I tried putting stuff in my bottle and it kept poking out of it too.
>>101717089>>101717050Oh it actually is in this one too. Bottom left has some green on the outside.
>>101717089I don't know anything about how this shit works; in the case of SD if I'm not mistaken it fundamentally operates in small square grid units which can have the downside of being overly discrete and "hard cutoff". Is there something about Flux that makes it the opposite, more blended/blurred units?Picrel did a bit better but the plastic of the TV is also a little thicker which might be way
>>101717071E1M1 will always be iconic
>>101717099Good catch. Realizing only know that this >>101717107just totally continues the outdoor part into the TV screen, so I take it back that it worked well.
>>101717118Here's the DALL-E 3 output I'm trying to replicate for reference
The SD damage control on Reddit was hilarious, comparing Flux to models like SD3 large and saying they are on par. Yeah, sure, Stability had a model as good as Dalle on their pockets, which is why we never got to try it kek.
https://github.com/bghira/SimpleTunerFlux.1Preliminary training support for Flux.1 is included:Low loss training using SD3 style loss calculationsLoRA or full tuning via DeepSpeed ZeROControlNet training is not yet supportedTrain either Schnell or Dev models
>>101717122>>101717118You might have luck changing your prompt a bit with something like "TV in front of brick wall" like in that Dall-e pic. Otherwise it might just focus on the outside city part of the prompt. Still messing around myself so I'm totally guessing right now.
hmm I changed the clip to fp16 and just the unet weight to fp8 and gens are working fineso I guess fp16 only makes things super slow on < 24GB cards.
>>101717136>the same retard who said flux was impossible to train has added flux training to his appThis is too funny
Will training be able to unfuck the lack of NSFW in the base model?
>>101717136>See hardware requirements.>A100-40G (LoRA, rank-16 or lower)>A100-80G (LoRA, up to rank-256)>3x A100-80G (Full tuning, DeepSpeed ZeRO 1)>1x A100-80G (Full tuning, DeepSpeed ZeRO 3))>Flux prefers being trained with multiple GPUs.its joever, no flood of generic character waifu loras for FLUX on civitai confirmed
meme magic. (but for real)
>>101717160Yes. It is very easy to gen naked people. It takes very little as it is to go from pimples to nipples.
>kigurumi still produces hoodie pajamas I don't get it
>>101717165Holdup, shouldn't it be possible to train loras once you have a a100? That doesn't cost too much once you know the kinks of lora training it shouldn't need 10 tries to bake one. The realm problem (like always) is actually making a good finetune of the whole model. That needs 3 cards and vastly more time. Is suspect the code needs to, and will, improve over time, but 3 a100 isn't all that crazy
>>101717174>change NBC to Bloomberg Newsit's so nice having an open source model with good text encoding now. All we need now is an inpainting tool so you can drag sdxl/ponyxl gens over and do edits involving text or logos easily.
>>101717190bro .. obscure fetish terms are not in, as much as I dislike it, but it also doesn't know shibari
>>101717209cool texture
>>101717165They cost a dollar an hour to rent, Anon. The bigger question is how long will it take to train.On SDXL I could train a LoRA in 20-40 minutes on my 4080, If it's comparable, or even twice as long, I can definitely see people spending a few bucks to make a LoRA. I will.
>>101717136>$55797 minimum to comfortably finetune Flux>$18599 for even a basic 64 rank LoRATell me that it's over without telling me that it's over.
>>101717215ya sure, but the coomer that wants to make 100 character loras wont have a100 equivalent cards for a while.. not sure if that is a blessing or a curse
>>101717228In a few years it will be possible
>>101717228>A100-40G (LoRA, rank-16 or lower)>A100-80G (LoRA, up to rank-256)>3x A100-80G (Full tuning, DeepSpeed ZeRO 1)>1x A100-80G (Full tuning, DeepSpeed ZeRO 3)
>>101717222It's about as much in as it was on 1.5 mergesMost of the time it would make weird suits, but you are right that it didn't try with a mask at allI've always been confused why it turns into sleeping wear at all
>>101717250
>>101717228Nigga literally WHERE THE FUCK do you get your informations from? Bait or retarded? A a100 cost like 7 $ per hour, lora training takes like 3hours at most.
>>101717241There are no scaling laws for GPUs. It's all stale and led by two corporations who engage in price-fixing.
>>101717228>>101717265>a100 cost like 7 $ per hourWrong. This is literally the first link on Google too, could probably get it cheaper if you shop around.
>>101717281>$1.60 an hour for an A100How exactly do they make a return on these things?
>>101717272AI is the first time since like early-mid 00s video games that upgrading to the next generation actually doubles your performance
what's the min vram for flux now?
>>101717281>>101717265>not using gold plated A100's
>>101717265(((He wants to rent a GPU like a good goy)))This /ldg/ not /cloud shit finetuning general/
>>101717293By not paying having to pay your own electricity bills
>>10171729612gb is a nice minimum
>>101717305i think he meant the people who rent out their own hardware
>>101717296I'm running it with 8. Apparently another anon ran it with 6. I can't say it's worth the wait though. I'm getting good results but the time it takes is 1~2 minutes on schnell (no idea why it varies so much) and 6+ on dev fp8.
>>101717318>>101717306bless
>>101717316Yes, they "make" money because their parents pay the bills
>>101717301If I had an A100 I would do it locally.
How do you even buy an A-series card? Do I have to be part of an institution in order for nVidia to be willing to sell them? I don't trust buying a $5000+ video card from third parties.
>>101717356No you can just buy them, but good luck. They are sold out everywhere.NVidia is the richest company for a reason.
>ask for an ugly person>all I get are good looking models that look like they are shot in a professional photo studioI hate this.
Whats utterly funny to me is that blackforest could absolutely easily train a godlike model thats way better artistically than midjourney, it has godlike prompt adherence + makes almost always perfect bodies. The craziest part is they all did this without ANY nsfw material. with 6.1 mj i get sometimes nipple slips. meanwhile flux has better anatomy despite having no nudity in the base model. Crazy. If black forest labs would want to they could make a mj level model effortlessly. This is what happens when you have the whole team that actually created the tech make the model.
>>101717144>You might have luck changing your prompt a bit with something like "TV in front of brick wall" like in that Dall-e pic. Otherwise it might just focus on the outside city part of the prompt.
>>101717356Just be like /mlp/ and steal them
>>101717301Calm down, Anon. If we want to finetune a model for local use we'll use whatever works for it. Send me four H100s if you disagree.
>>101717228https://www.ebay.com/itm/276575178482SXM4 prices are a little better but it's still a horrible dealWhen will some autist reverse engineer VRAM and just allow us to add these as we add RAM? I know some modders have doubled it but what else can be done?
>>101717370wtf is this post
>>101717379Is this how Pony Diffusion became a thing?
>>101717379>let's just believe random shitpostswould be better with an image of the HW
>>101717369skill issue
>>101717391Doubt it. Last time I checked, mlp was full of dalle plebs who don't know shit about ai.
>>101717410That's just a normal person. Make me a freak junkie character like in that mad max movie.
So you can train a rank 16 Lora on 2 3090s?
>>101716019>You dropped this
>>101717391Ponys are suspiciously wealthy, like to an ungodly level. Before diffusion they would pay thousands for porn commissions from real artists. They put their wealth into training PONY XL and we all reaped the rewards.That's the real reason artists hate AI.
>>101717433How the fuck are you guys pooling vram for lora training? Thought this was impossible
>>101717373Ok this one is the closest but it absolutely refuses to draw scanlines or any other kind of declarifying style
>>101717425
>>1017174641980's televisionSecurity camera footageJust throwing out ideas.
>>101717449>muh wealthy furfagsFirst off, that's only a few of them and the ones that spend ridiculous amounts of their fetish are also the loudest ones, so they'll be over-represented.Secondly, you underestimate how easy it is to scrounge up money for your fetish if you simply cut down on everything else in your life.
>>101717425>>101717465
>>101717433Doesn't fit into 24GB so you can't train
>>101717370Sure, but the fact is that they won't train with nsfw, so this model will just be at most edgy nazi trump whatever memes.I wonder what is in the pro model though, if it's actually less censored or it's just better.
>>101717483prompt?
>>101717479Sure, but they fund their fetish while being a minority, way more than the vast majority of people into human looking characters, despite being way larger of a pool.
>>101717479So what you are saying is you are both a pony and poor>>101717500>low quality snapchat photo circa 2015 taken on an iphone of a meth addicted ugly man with asymmetrical face, large nose, missing yellow teeth, big ears, bad skin. He is a meth head, with a gaunt face, and is extremely ugly, he has sores on his face and sweats a lot,
>>101717485You can't train across two gpus?
>>101717370>If black forest labs would want to they could make a mj level model effortlessly.But why would you want that? MJ is MJ. And it's terrible imo. They tackled Dalle and they did so nicely, by tackling Dalle they have in turn also tackled MJ because it truly isn't asd capable if you can simply prompt for everything that MJ does yourself (whereas MJ only does it without you asking, hence creating the misconceptions that it's better). Flux pro is capable of really aesthetic gens as well, everything posted to https://blackforestlabs.ai/is aesthetic.
>>101717521you can't pool vram for it, no. multiple card training just speeds up the process but it still has to fit on the vram limitations of one card or you oom Nvidia used to offer a way around this by linking your multi cards together to share vram but in true Jewish fashion they removed that ability from any modern card
>>101717533mj has a very specific mj look. you can always tell when an image is mj, and that's what I don't like about it. I want us to get to the point where we can't tell an image is ai at all. Flux is a step closer.
>>101717547You mean nvlink?It's still there in the 3090, so you can at least link 2-4 cards.
>>101717558Not the case with MJ anymore (at least not for Niji or realism prompts). You can usually pick apart the art pretty easily though. The main issue with MJ is lack of control imo.
>>101717547>they removed that ability from any modern cardwrong, they removed it from consumer cards, workstation cards like A100 still have NVlink capability
>>101717599I love paying more for less
>>101717547I was sure 3090 doing that was like its one feature that it had over 4090.
>>101717622exactly, 3090 (and other 30xx series cards) still have NVlink for 40xx series NVidia removed that for consumer cards in the specs theoretically a custom version of the 4090 could do nvlink, but afaik there never was one released by any manufacturer
>>101717622I have 2 cards and I can nvlink them easily since I know where to get one, but there is no point since no software seems to take advantage of it right now.
>>101716890It does work but only limited to Schnell fp8 model. That means you can't gen meme text image unless you wouldn't mind to play gacha or wait for 20 minutes for a 1024x1024 20 steps on fp8 dev model.
how is 2048x2048 on flux?
>>101717707meh, if I had that VRAM I could tell ya, but I can tell you how it behaves up to 1536x1536: It depends on what you prompt but depending on subject you a rasterization effect appears like in >>101717209 (yea was my picture but thought the effect looked cool on that) that stretches with the resolution the higher you, more visible it is, this is most present when prompting painted or illustrated stuff like anime, old paintings, etc. .. for realism it isnt as extreme but it is visible at a certain point. My guess is that the data that was given in 2MP was mostly photo realism, while other subjects were in 1MP or less.. using more steps somewhat mitigates the effect but its still visible if you zoom in, my bet is the optimal range to prompt is 1MP-1.5MP, above is pretty meh
>>101717707Functionally impossible to gen for me, so I couldn't tell you. 1024x1024 is good enough. Can always upscale in a cheaper model.
>>101717165>>101717265>>101717228Niggers. 240GB rigs cost peanuts at vast.ai and similar places.12B transformers are routinely trained on these, check how many trash checkpoints there is at HF. Renting it not expensive.> flood of generic character waifu lorasThankfully it's unnecessary for a model like this (although possible), you need pony-grade finetunes
>>101717660Right, so the issue is not the models, but the software like comfyui (or whatever libraries it uses)?Like it COULD be done?
>>101717899https://huggingface.co/TheBlokeThis guy had quants of nearly 4000 finetunes of LLMs of sizes up to 120B. There are people who train checkpoint after checkpoint of 70B models only to throw those out. All of that while faggots at /ldg/ still live in 2021 and whine about mad bucks required for this.
>>101717964what happened to the guy?
>>101717973They got him. (he started a company and has no time/desire anymore)new useless LLM finetunes are still out every fucking day
>>101717973>https://huggingface.co/TheBlokeprolly hired by big tech
I wouldn't worry about finetuning costs. Dataset preprocessing cost (VLM captioning) dominates anyway, this is true for both SDXL and Flux.
is flux still broken on mac? my gpu crapped out suddenly stopped working and I had to switch back to my m1
>>101718053LoRAs will be fine to start with. Can get some style LoRAs easily enough. I don't mind renting some GPUs to get them done, but I want someone else to bite the bullet and do the hard work for me and figure out how to train it.Anyway, full fine tunes are not my problem, someone else will do them, and I will consoom.
>>101718053>Dataset preprocessing cost (VLM captioning) dominates anywayHow many % of total finetune costs is captioning?>>101718106>LoRAs will be fine to start withI have 200-300 captioned datasets ready
>>101718089do you have a mac with 64GB of unified RAM? cause thats about as much as you need, else you are outta luck
Is this real or flux ??https://mobile.x.com/Rainmaker1973/status/1819976401618903451
>>101718158dead internet theoryhttps://mobile.x.com/fopminui/status/1819977157814210723
>>101717902Probably.But having multiple cards + nvlink seems rare.Most people have one card.And the ones having multiple cards don't have or can't have nvlink.
>>101715949>Avoid ldg for a few months because AI art started to look too samey >Decide to check it out again>A new model gets introduced>Actually looks really fucking good>Can't run it FUCK!
>>101718190ACKJust when I was starting to think buying a 4090 was a CONSOOMER mistake... :3
>>101718158
Can you run Flux fp16 split across multiple gpus? I’ve got a 3090 and 3060, would be nice to have the text encoder on the 3060 with the main diffusion model on the 3090
>>101718194right? I am so glad I got one right now
>>101718158prompting the image description on flux kek
>>101718204I am sad I cheaped out and spent $1500 dollarydoos less on the 4080, but she's been good to me, and I can run the FP8 and it's producing exceptional results. SAI have to be shitting their pants right about now.
>>101718203Yes, use : https://gist.github.com/city96/and add the picrel nodes.This cut my gen time by 50%, basically the 3090 never swaps ram from what I'm seeing in htop.
>>101718216
>>101718169Book text looks too well written to be generated.So this looks more like the usual clickbait accounts answering each others randomly.
>>101718237Nice, thanks
>>101718199nothing wrong here, it's a ramp. there is another ramp but made of cement ahead of the girl.
>>101718240Reminds me of the African kid water bottle FB pics.
>>101718295My son made a fully functioning exoskeleton out of trash 1 like = 1 prayer
>>101718234Give booba prompt
>>101718237It doesn't seem to do anything for me, if anything makes it 3 times slower for some reason, incredibly pissed cuz no idea how would i even debug this
>>101718194It is. Used 3090 goes for $700 (may have gone slightly up since flux though)
>>101718374a photograph of 2 women laying on the grass. The woman on the left has small breasts, and a white tshirt that reads "2B is all you need" and the woman on the right has extremely large breasts and cleavage wearing a tshirt that says "Flux 12B"
>>101718382
>>101718412>extremely large breastsoh...
>>101718421nta but do you think if I steal my sons 2060 and shove that into my pc I can do fp16 on my 4080?
>>101718237Oh, so to go 50% faster I just add my 3060 ti 8GB which has been eating dust?
>>101718382It's pretty much fire and forget anon, don't know what's going on in your workflow, share it maybe?For me, without : >Prompt executed in 51.17 secondsand with it : >Prompt executed in 25.22 seconds
>>101718428prompt for large breasts weren't large enough
the t5xxl & clip_l guidance prompts being allowed to be separated is ocd inducing- which parts of prompt work best where, same seed, such varied results.. fcuuck
>>101718249add gaming chair
>>101718439I have 2x3090s so I can add both the clip and vae in the second card to get it done faster, but I think the second card goes up to 9GB vram (see picrel), so maybe try adding both nodes, then just one of them to test.
>>101718502If it's a 3090 I probably won't get near the same speed improvement even if I can fit it into the card
>>101718532/lmg/
>>101718543Not but it'll be better for sure.The less you move the model between vram and ram the faster.Worth a try.
>>101718477Not only that altering the CFG basically improves or degrades image quality, so if you think it sucks going from 3.5 to 2 might help etc...
>>101718549>>101718568I agree but please understand that BFL is SD. Flux is SD. They're the main people who developed latent diffusion, and the ones behind SD1.5 and XL. SAI has no talent left.
>>101718613>SAI has no talent left.Who the fuck keeps pumping millions of dollars into them?
Prompt: a manly anime version of barack obama wearing a two piece women's bathing suit, with massive enormous radical puffy supple female breasts
>>101718620Actual retards who don't even do simple checks on the people they piss money into. SD has name recognition and that's all they are coaxing on.If 3.1 isn't as good as Flux on half the VRAM SAI are done.
>>101718640>If 3.1 isn't as good as Flux on half the VRAM SAI are done.80% as good but with nudity/nsfw possible would be awesome.
>>101718640SD3 should have been the final nail in the coffin. Anyone with any eyes on the company would realize they had been pumping out shit and stalling for time for a while now. Between every image generation release they'd make like 50 LLM finetunes or audio models that sounded like screeching from hell and nothing of substance.
I just want to say that I am exploring the latent space and it's genning a lot of pictures of children.
>>101718650
https://www.youtube.com/watch?v=Ila5tQ6vCBoLooks pretty damn neat
>>101718656And yet somehow, even after that trainshrek of a model, they managed to secure another 50m.I don't understand california.
>>101718706perhaps money laundering
>>101718686Can't wait for future models to only allow cat memes.
>>101718704I use this sometimes for simple objects. It's good, especially for building modular parts to buildings but man is it finicky. Like it controls like ass
>>101718662Cute but how do you get rid of that always on depth of field?
>>101718718I kind of assume it was money laundering, considering how much they were eating through each month with literally nothing to show for it.
>>101718724I wonder if it would serve as separate inpainting program for normal 2d images
I was told that PonyXL is not fine tuned on top of SDXL but is a base model with the same architecture, is this true? If not, where can I read the announcement/explanation about how it was trained so it'll say that it was finetuned on top of base SDXL?
what are your samplers and schedulers of choice for FLUX? does a comparison grid exist somewhere?
>>101718751If you want to project the image on a 2D plane, sure, but there are so many better options out there for that.
>>101718613When OpenAI talent left to found Anthropic, Claude didn't suddently get called ChatGPT. You are coping.
>>101718731
>>101718777>but there are so many better options out there for that.I prefer not using comfy. I think ReForged has borked inpaint + controlnet at the moment
>>101718779Because ChatGPT isn't dogshit. Flux is the next step in Stable Diffusion. It's what SAI wishes SD3 was.
>>101718731you dont because it has no negative prompt and was trained on ideogram slop that has that style and dof built in by default
>>101718830>Because ChatGPT isn't dogshitKek
>>101718838astute. I canceled ideogram, because this is better.
>>101718861>I ignore reality because I like to be contrarian and hate on the popular thingopinion discarded
>>101718867Same. My flux gens are much better at text and design than anything I genned with ideogram and their model (at least for humans) appears to be stuck in SDXL era when every other model has moven on.
>>1017188804o doesn't hold a candle to 3.5 Sonnet anon. GPT stopped being relevant past 4. Turbo was absolute flop in comparison to Opus and they have been flopping ever since.
>>101718830do you have it? Ask it which side the gas cap is on the Nissan Sentra.
>>101718934I develop AI tools for my work. I use Claude in those tools, because it's contained within AWS and our data is secure. GPT is still better.
>>1017184453080ti - best i could do is prompt executed in 92.28always getting the lowvram message no matter what 10041.5125idk if images keep their workflow when uploaded here
>>101718934For rp.For assistant chatbot gpt is hugely popular.Thus the deluge of "it's important to note" in everything, from emails I receive to random posting.
>>101718950I don't even know that nigga
>>101718957You are a liar because I also use it in my work and 4o hallucinates and straight up sucks (though the smart way is to use both, but yeah 4o is most useless for what I do).
>>101718982>claude never halucinatesWe're done
>>101718982why use 4o when Turbo is right there?
>>101718962You need to upload your workflow to catbox, 4chan strips everything.Also, a 3080ti? And just one card? Why do you use these nodes if you just have one card?
You HAVE protected democracy today, right Anon?
>>101718934>>101718982Isn't 4o incredibly cheap though?
>>101718982Lower the temp.
>>101718934Why was the free chatgpt quite clever, and then very stupid, so fast? I was using the free one, and it was nice, and then it went stupid and literally useless.
>>101718972It's supposed to know stuff we don't know.
>>101719014forgot to pic
>>101719034No it isn't. It's a predictive text model. Everything it does is guessing.
>ywn have a patriotic woman with a belly riflewhy even live
Bigma status? Keeping my potato warm by the fire in the meantime
>>101718970In my case it's useless even for assistance. >>101718998I want something that is cheaper, not lazy/brain dead like Turbo.>>101719022Censorship.
>>101719118
>>101719135Skill issue.
>>1017190014 of them (1 currently down due to a riser)https://files.catbox.moe/tjuj7i.json
>>101719175forgot to mention but Im also using the comfyui install from swarm, idk if that would affect something like this
>>101719179Good night, cunt.
>>101719022Not sure if clever is the right description.I cant put it into words well but at the beginning chatgpt was about pleasing the user.Giving you something you didnt explitly prompt or realize yourself you want yet.Like reading between the lines, trying to figure out the purpose of the chat and then delivering. And i dont mean just coom, in general.Its never been the same. That was gone fast.Alignment means teaching the AI to be unhelpful to the user. The implications are severe.
flux's idea of a swastika, so I guess it wasn't trained on swastikas. India weeps.prompt: swastika
>>101719205someone got a swastika just fine yesterdayhttps://desuarchive.org/g/thread/101701058/#101701913
>>101719205???
>>101719118Someone posted a bigma demo yesterday, it didn't look very promising.
>>101719225kek, that someone was me.
>>101719260nice, can I get a catbox so I can get my own nazi catgirls?
>>101719282https://files.catbox.moe/n4g11y.png
>>101719304thank you, gonna see how slow flux is on my rig later after I download it all
>>101719282>>101719304By the way my flow sucks I was figuring out how to change guidance without realising there's a node for that.
>>101719258>Someone posted a bigma demo yesterday, it didn't look very promising.? It's not close to out
>>101719175Looks ok to me, did you try using just cuda1 and just the vae node or just the clip node?Otherwise can't help much, sorry anon.
is forge obsolete now?
>>101719366>is forge obsolete now?Nope, but it's more experimental now. Reforge is decent and A1111 with latest update seems as fast
>>101719418the update broke a ton of shit like hires fix
>>101719418thankswill look into it
I wish Flux knew celebrities like DALL-E 3.
>>101719435yeah if you wanna use base forge you better not update
Why after having caught up do we still get mogged so hard by Dalle threads? I guess the total brain death from SD prompting stifles creativity.
>>101719418reforge and a1111 are alternatives to comfyui?
>>101719456https://github.com/TencentARC/PhotoMakermaybe this?
>>101719467yeah why not
>>101719466I bet Dalle is juicing prompts with ai rewriting.
guys, what am i doing wrong hereit is taking me half a century to generate an image with flux3070, i510600k, 32gb rammy ram is being flooded while my gpu is slowly bouncing between zero and 70-90% workloadsi downloaded the Kijai flux-fp8 one from https://huggingface.co/Kijai/flux-fp8, should i just be using the one in the OP instead or did I mess up in the workflow settings somewhere?me big retard, plz help
>>101719466that's just you man. go back to your cope thread
>>101719489no like I am asking. I literally only paid much attention like 3 weeks ago and finally got things working alright.
>>101719485it's just not the same compared to being trained on them from the start>>101719501that's explicitly stated as something they do, even in the API, it's not a secret.
>>101719502use comfy's
>>101719316It was this >>101705795, no idea if it's legit or not, but some anons tested it and in its current form it seems pretty bad.
>>101718764mainly euler beta or euler simple cfg at 1... changing cfg increases step time and causes deformity or blown out imagesmost tweakable setting seems like guidance from 0 to 100still trying to figure out if splitting the prompts is useful for clip and t5, been putting style info into clip
>>101719515Ah ok. A1111 is for 1.5, sdx and sd3 at the moment (and pixart?). Maker of A1111 isn't buttbuddies with the tech trannies so you'll have to wait some extra before A1111 support for new things. You might as well have both installed.>>101719538>some anons tested itIt was complete garbage.
>>101719538That's lumina, but yeah. Pixart team started out pretty small with limited resources then they got compute from Nvidia. But in terms of talent they tend to be extremely frugal and haven't released anything quite as good as big models (Pixart was good but heavily undertrained). Expecting a quality model from them, especially on the caliber of Flux is nothing more than a wet dream. Hunyuan has a fighting chance because Tencent is a massive corporation but since they are not a startup there are valid concerns of censorship on their model. Though Tencent can still release tools that improve prompt adherence , future versions, etc... There is still a chance with Pixart, but it's unlikely it'll be as good.
>>101719567>guidance from 0 to 100what? it's completely cooked at 5 for me.
>>101719567deis ddim_uniform is my personal favoriteI have yet to play with the guidance, kept it at 2-2.5 mostly
>>101719538that's lumina, not pixart. different model from a different group.
Sprechen Sie Pomptisch?
how do I make her have a flat chest without prompting for a teenager?
>>101719806wait for PonyFlux
>>101716770I have fp8, what text encoder is being used? I didn't know about text
>>101719778Blitzgriegbesamung
>>101719806add "trans"
>>101719839I want that to be a thing
>>101719684i wouldnt call it completely cooked. it just doesnt follow your prompt as closely
>>101719567>>101718764what does sampler do, and scheduler?
>>101720035>>101720035
>>101719418forge reforge is still better than a1111 for VRAM usage
>>101716450cogv seems to have 8 versions, and no explanation, what are they for?https://huggingface.co/THUDM/cogvlm2-llama3-chat-19B/tree/main
>>10172014548gb of vram? I only have 16gb of vram, 32gb of system ram.
>>101720122that's not "8 versions" that's one checkpoint split into 5GB chunks so it needs just under 40GB to full load into RAM/VRAM
>>101720170>>101720169
>>101718234For that pic to be accurate, the 2B woman should be a deformed mutant half buried in the grass with the text on the t-shirt wrong
If anyone is still hanging around, there's a new thread up already>>101720035>>101720035>>101720035
