Discussion of free and open source text-to-image modelsPrevious /ldg/ bread: >>101940241>Beginner UIEasyDiffusion: easydiffusion.github.ioFooocus: github.com/lllyasviel/fooocusMetastable: metastable.studio>Advanced UIAutomatic1111: github.com/automatic1111/stable-diffusion-webuiComfyUI: github.com/comfyanonymous/ComfyUIForge: github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: github.com/invoke-ai/InvokeAISD.Next: github.com/vladmandic/automaticSwarmUI: github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outrentry.org/sdvae>Model Rankingimgsys.org/rankings>Models, LoRAs & trainingcivitai.comhuggingface.coaitracker.artgithub.com/Nerogar/OneTrainergithub.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhuggingface.co/spaces/black-forest-labs/FLUX.1-schnellcomfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThuggingface.co/spaces/PixArt-alpha/PixArt-Sigmahuggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThuggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: github.com/city96/ComfyUI_ExtraModels>Index of guides and other toolsrentry.org/sdg-linkrentry.org/rentrysd>GPU performancevladmandic.github.io/sd-extension-system-info/pages/benchmark.htmldocs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: www.mage.spaceimg2img: huggingface.co/spaces/huggingface/diffuse-the-restsd3: huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>Maintain thread qualityrentry.org/debo>Related boards>>>/g/sdg>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
are loras working with GGUF without tanking the performance yet?
>mfw
>>101929139Neat!
>>101943981amazing font too
>>101943957yeah all of them in forge
how would you go about making a prompt for a POV shot that shows the POV's hands and other body parts (I mean legs)
>>101944048"sees * from a first person point of view"
I had to manually draw the Italian flag colors patches for these then inpaint over them at 0.2 denoising because the LORA just would not get them even close.
On this one, the original gen had a short sleeve shirt but the rest was so good I kept it, then successfully inpainted over the bare arm, set to inpaint whole picture, at 0.9 denoising with long_sleeves tag and short_sleeves negative, with Canny set to 0.1 to keep the position of the arm, then did a little shooping and inpainted over again at 0.2 denoising with it set to inpaint masked only to fix the slight fuckups introduced by the first inpaint.
Since she's no longer wearing pantyhose here, I like that the model is smart enough to adjust the skirt position when I negative prompt pussy and panties, instead of just doing barbie doll anatomy.
Blessed thread of frenship
>>101944196okay but surely you made a version that isn't censored
>>101943963hello!
>>101944218Why would I do something like that? What a preposterous assumption. Absolute poppycock.
>>101944260catbox now
>>101944231>reflectionNice!
>>101944253what is it
>>101944306Probably One Punch Man artist lora
>>101944311rad
Any tips on prompting for Flux?
>>101944314https://civitai.com/models/655094/fluxdev-yusuke-murata-art-style-lora?modelVersionId=732897
>>101944288Wait, I take this back.>reflection doesn't draw the chimneyAbsolutely disgusting.
>>101944343Thanks yeah saw that earlier.Btw how did you get that clothign rip near the navel, was it chance or a prompt?
>>101944344cant win them all
where can i use flux for free i cant run it on my computer and everything is too expensive because hyperinflation in my country
>>101944366WHY THE FUCK DOES SHE NOT HAVE A CHOKER!?
>>101944384ikr
>>101944380https://huggingface.co/spaces/black-forest-labs/FLUX.1-devtwo uses per IP, enjoy
how do i run flux dev nf4 on forge? it's not recognizing the model typei know consensus is that q4 is better but i want to compare them myself
>>101944458>nf4Why would you do that? There are better versions of the same size
>>101944465This is an unrealistic image because nobody wants to have sex with, well, you know.
>>101944480he literally answered why in that post retard
>>101944465I’d breed her.
>>101944488plenty of oil drillers out there
>>101944530Oilers is my favorite team
wait, pictures created using Flux cant be used commercially? wtf?
>>101944566>furpuck
>>101944582kek, how they gonna know
>>101944582What are they gonna do, sue me?
>>101944582that applies to dev model only, not schnell
>>101944598super secret watermarks embedded into every image and impossible to detect or remove
>>101944598it's about catching large scale operations, no one actually cares what some NEET with $20 in his checking does
>>101944582And how would anyone know if that's a Flux dev images I'm selling, huh?>"b-but there's an invisible watermar-"Oh nooo I just accidentally inpainted it with schnell at 0.05 denoise, whatever will they do now???
>>101944617>impossible to detectIs this possible by placing it in image data only and not metadata?
>>101944635nobody knows
>>101944582just lie and say you made it with schnell instead of dev. burden of proof is on the accuser>>101944617meds
>>101944635You could theoretically embed in the pixels some sort of watermark pattern. But it's unlikely because that would be obvious in the code. At the end of the day, it's to stop people from putting Dev on a commercial API.
>>101944649That's wrong, in a civil suit you would have discovery in which you would need to prove you were using Schnell and "trust me bro" isn't really a good defense.
>>101944668>in which you would need to prove you were using Schnellthe burden is on the prosecution to make the case, not the defense
>>101944690>we have money and they don'tprosecution rests
>>101944690Actually it's not, this is a civil suit, not criminal, the burden of proof is much lower and judges can off-handedly determine your guilt especially when you refuse to show that you didn't use Dev.
>>101944708>civil suit>against someone with no moneyThis doesn't happen.
>>101944727they would only sue you if your commercial use was making money
>>101944727Then you have nothing to worry about. The enforcement of the commercial agreement would only happen if/when you were making money.
>tfw you're getting sued over a 1girl, spread anus, hairy butthole picwwyd?
>refuse to showi didn't save the prompts your honor. why would I? i was never doing anything wrong
>gen time inexplicably risen by 30%hmmm.
>>101944744>>101944754>it's making $10 per dayThey're not gonna sue. They're not gonna sue if the defendant doesn't have the funds/insurance to even pay for the lawyers to file/litigate the lawsuit.
>>101944772Dog ate my homework isn't a valid defense.
to be honest I still like my SD 1.5 gens... But I like how it always generates correct hands in Flux. Has anyone thought of a workflow that uses Flux only to fix hands?
>>101944790They will sue if the value of the chilling effect is worth it. $10k to shut you up and garnish your wages for the next 10 years to discourage other people goes a long way. Frankly, you're retarded if you willfully try to make money with Dev. Frankly, if you were intending to make money you'd spend the $1000 to full fine tune Schnell to make something that is both exclusive to you and doesn't have the risk of being fucked.
>>101944790you're too young to remember when the record labels sued people for downloading music off napster
>>101944762Highfive the judge who is a based boomer who knows I dindunuffinfs and generate some more
>>1019448231.5 is so fucking nuts, you can create entire art styles just from how much negative presets break the models. I created this weird style in 1.5 ghostmix using a combination of several negatives i can't recreate in pony whatsoever.would love to try to niggerrig train something at least close in Pony but i dont have the time to dedicate with this slow gpu.
>>101944762>your honor, dev can't gen buttholes THAT hairy, it's definitely a ponyslop pic
>>101944837I am old enough and I do remember that. I also remember it was an extortion scheme. They didn't actually intend on taking downloaders to court in most cases - just hoping people were stupid enough to pay out of court because they were scared.
>>101944920fuck around and find out then, retardI'm sure the cloud provider you host your Dev endpoint on won't keep any logs showing which models you used
>>101944864OBJECTION!https://civitai.com/images/24567282
>>101944582All this argument over something that isn't even true because you retards can't even be bothered to actually read the short license>Outputs. We claim no ownership rights in and to the Outputs. You are solely responsible for the Outputs you generate and their subsequent uses in accordance with this License. You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein. You may not use the Output to train, fine-tune or distill a model that is competitive with the FLUX.1 [dev] Model.>You may use Output for any purpose (including for commercial purposes)
which quant do i run with 16GB VRAM and about a 3080's worth of compute throughput
>>101944690civil suit, discovery, etcbest be ready to flee to venezuela
People are LTERALLY making money right now by releasing early access loras on civitai
>>101944582only models, gens have no licensees
>>101945059It would be ultimately Civit's liability
this pic goes hard
>>101944145
No one asked for this, but reporting.Sold my 7900XT and received an A4500. Under ROCm flux ran ~2 s/it. Same prompt and workflow, A4500 flux ran ~1.4 s/it
>>101945174nice
>>101945124Are you using this?https://civitai.com/models/638000/arnold-schwarzenegger-1990s-flux-lora
>>101945159>posting a picture of an incorrectly held handgun on an anime website
whats that node setup for ultimate SD upscale? someone mentioned a few threads ago some setup for calculating the necessary settings per latent image res just tried to do this one stocking gen and it ended up fucking up a few details, presumably because i kept my 1.3 scale settings for 2x kek
>>101945197yes
>>101945174surprisingly, using a bigger model (q8) on the same seed and more steps (25) degrades the quality for some reason
>>101945273>>101945188Claude helps:>The art is in a style reminiscent of Egon Schiele, the Art Nouveau movement, and Pop Art movements, representing a dialectic of vibrant colors, dynamic brush strokes, and detailed textures.
can Forge pin T5 to the cpu yet?are VRAMlets just waiting for it to move T5 around every time you change the prompt?
>>101945298vramlets are unironically better off using comfy. Not only it's possible to offload anything anywhere, but there is also a much more optimized lora support for gguf
>>101945324>>101945343what are these guys up to
>>101945340>but there is also a much more optimized lora support for gguflast time I tried using loras with GGUF in Comfy it doubled the gen speed and there haven't been any new commits
>>101945340former forge user here on a gtx 1080, it works better solely for one specific point:Ultimate SD Upscaleri just scaled an image to over 2k pixels while my vram usage stayed under 7 gigs the entire time. Unbeatable. Also i need to learn how to offload the text encoder to the cpu for an extra boost been meaning to look into that again, but there's just SO MUCH SHIT TO LOOK INTO. Like right now im looking into this so i can optimize scaling even further >>101945262
I'll be honest. I don't like Murata Yusuke's art style that much. It's fine for One Punch Man but outside of that, it doesn't feel great.
>>101945355Perfectly normal, non-summoning demons, stuff.
>>101945262Just get your input image's width and height using a custom node and feed them into your upscaler node (right click - convert input). Generally the default tile padding and feathering work just fine, you might want to set tile order to chess because it (supposedly) helps counter visible tile seams. Denoising strength is one thing you have to play with.
>>101945262you just route everything into it and turn the tile size into inputs and then, well you should use a tile size the model was trained on, I just route the original gen into a "get image size" node and use those numbers. add some mask blur and no seam fix. pretty straight forward - for flux, try 16 ish steps and start at 0.2 denoise
Clean it up janny
hello, i was told you niggas had jailbroken joycaption
>>101945546it describes nudity and sex out of the box
>>101945370I'd take it over forge "patching" shit into your model every time you hit generate
>>101945557that's why i was curious
>>101945574even when you don't change any lora settings? cause that happens with comfy but only when you add/remove a lora or change the strength
>Natural languageA table with a fruit bowl on it>SD 1.5Table, fruit, bowl>Flux """""natural language"""""This is a photograph of a ceramic bowl filled with various fruits, placed on a wooden dining table. The bowl, which has a ribbed texture and a glossy white finish, is positioned centrally on the table. Inside the bowl are several pieces of fruit: a bunch of ripe, yellow bananas with a few brown spots, a couple of red apples, a peach, a plum, and a small blueberry. The bananas are slightly curved, and the apples have a smooth, shiny skin. The peach has a vibrant orange-red hue with a hint of yellow, and the plum is a deep red color. The blueberry is small and dark blue. The wooden dining table has a natural, light brown finish with visible grain patterns, giving it a rustic look. Behind the table, there is a wooden chair with a simple design and a light brown color, matching the table. The background features a wall painted in a light green color with a subtle vertical pattern. The overall setting suggests a cozy, homely atmosphere, possibly in a kitchen or dining room. The photograph captures the textures and colors vividly, emphasizing the contrast between the smooth, glossy fruits and the rough, natural wood surfaces.
>>101945433>>101945468thanks ill take a look at that in a bitprobably, honestly it looks like i figured this out well enough, got the highest resolution i could realistically want, good performance for my hardware.haha lol love this fucking lorahttps://civitai.com/models/610167?modelVersionId=682029https://files.catbox.moe/31ixoc.png
>>101945637>A table with a fruit bowl on itWhy do people insist on lying? You don't have to boomer prompt.
>>101945637>noooo i don't like being able to specify the details of exactly what i want
>>101945657>i just noticed the pizza morphing out of the boxi just felt one of my nuts shoot into my chest, god dammitthe nude ones ended up better kekhttps://files.catbox.moe/3qr8b9.png
>>101945637lol wutI guess you could prompt like that if you're retarded
>>101945637Could do better.
>>101945691Now this is prompting
>>101945691show the gen
>>101945691run it
>>101945691Fuck it, do it live.
>>101945706>>101945709>>10194571820%| | 4/20 [01:22<05:18, 19.93s/it]
A picture is worth a thousand words, as they say :)
>>101945691Do we have "max" token cutoff?
>>101945720kek
>>101945726I heard it's 512 tokens
>>101945720still faster than when i tried to run flux nf4 on my 1080 lmao
>>101945657glad to help & nice gen lol. the important thing is to raise mask blur to like 16 or so (or even higher; should also increase tile_padding to at least 64), the seam fix is BS.
>>101945706>>101945709>>101945718>>101945720Here the result.
>>101945658Flux slopa>>101945791Flux arta
>>101945765yeah one guy mentioned mask blur 8-64 i think. Ive been keeping it at 16. What exactly could change if i shot it up all the way to 64?As for denoising, I was also suggested to try 0.05, that.. fucked my gens hard. I'm using 0.15 and it seems fine judging by those gnes.>>101945791L-La creatividad..
>>101945791VERY NICE BRO. really very.. real. dangIT BETTER BE
>>101945288
>>101945807You don't have to boomer prompt.
>>101945791those blueberries are massive
>>101945842You'll never be a real AI artist.
>>101945791You made me read your prompt novel to check; but that's some pretty good adherence IMO.
>>101945726This is dev 30 steps?
>>101945853Also, it's with Q4_0, probably would have been better at full weight or Q8_0.
>>101945791Goddamn that's nice bowl of fruit
>>101945809it is really just applying a blur effect where the tiles meet, and that value is in pixels. 16-32 is enough. 0.05 denoise isn't. (if we are talking flux here) 0.2 is a good ballpark. step wise, I found that below 16 you loose detail but higher not really worth it
>>101945791promptlets on suicide watch
>>101945860This is one of the first ones I did with Flux... so, yeah. fp16, euler, simple, 20 steps and 1.8.
>>101945887Oooooh gotchaI have to say i'm blown away how flexible this system is. Didn't think it could be possible to upscale on this gpu and have so much vram headroom.And no not even gonna try running flux until more LORA's come out and 2D gets more progress.
we catfishing with this one boys
>>101945952https://www.reddit.com/r/truerateme/
>>101945952Nice brush strokes, lady.
>>101945952saw her in a J horror
>>101945927Looks good, did you use anything special in the prompt, natural language wall of text?
>>101945842You can let ollama boomer prompt for you:https://github.com/wujm424606/ComfyUi-Ollama-YNhttps://github.com/MinusZoneAI/ComfyUI-Prompt-MZ
>>101945952>No buttchinYep she's real fellas
Hi
>>101945952make it a thin pen writing and give her acne and make her melancholy and then you're starting to have something truly catfishable
>>101945998I just used an LLM with that prompt:Help me make a prompt for an image generation model. My original prompt is:```original prompt here```I need something more verbose to have a better output.
Help me make a prompt for an image generation model. My original prompt is:```original prompt here```I need something more verbose to have a better output.
>>101945964maybe I should work on getting a super one just right then make one of those>EVERYONE SAYS IM UGLY :(((( AM I?posts
>>101945993>Capture this image using a high-resolution digital SLR camera with a wide aperture (f/1.8 to f/2.8). Set the ISO between 800-1600 to balance performance with minimal noise. Focus on ultra-sharp facial details, rendering each pore, fine tan lines, and subtle skin variation with meticulous precision. Capture individual strands of hair with distinct clarity, including flyaways and the subtle sheen of her gloss. Render the eyes with exceptional detail, showcasing the reflection of light, and the moisture on the surface.You probably don't need the vomit now that we have a dedicated lora.
>>101946039same thing, just integrated into comfyui so you can turn it on / off and use your prompt box as usualcan also be used to translate to chinese for hunyuan or comparable
I want to casually screw around with flux. Huggingface free generations ran out quickly. I could install locally but I only have 8GB VRAM GPU, so slow generations right? What's my best option? Any other online instances of it that aren't shit?
I have been out of the loop for a few months. What is the best current model for anime, sd3 flux, sdxl? I don't care about licenses.
>>101946121pretty sure anything 3000 and 4000s can handle flux fine enough on 8gb of vramolder gpus take forever
>>101946081That's a good prompt, I used a similar one but I think this one does better with flux, it really did the hair good
>>101946123PonyXL
>>101946123pony
>>101946123PonyDonutHoleXL
>>101946121People came up with tricks like gguf quantizations and so on https://civitai.com/models/647237These may not be too bad.
>>101946123PonyBigStinkyAssholeXL
>>101946123Autismmix Confetti, which is a pony derivative.Don't use OG Pony, it's vastly inferior.
>wake up today>gens are two times slowerCOMFYYYYYYY
>>101946123>bestnovelai v3>localautism mix/pony
Have a bootleg Nicki Minaj
I'm cumshot Flux lora guy from earlier BTW, go here to this /b/ degenai comment if you wanna see a monster banging a chick as generated by Flux (and also some of my thoughts on how things might go for NSFW for Flux):>>>/b/923027188
>>101946134Yeah I should have specified I have a 1070. So no local for me :(
>>101946123Pony is best on average,Flux has advantages with (currently mostly SFW) interactions between characters and complex prompts.
There's quite a lot of anime ponies I prefer to any version of Autism nowadays personally
fug didn't see the new bread>impatiently waiting for JoyCaption to batch tag 10k+ pics before i can start lora training
>>101946121>>101946195>Any other online instances of it that aren't shit?https://fluxpro.art is your only other hope
>>101946194yeah no
>>101946187Autismmix is way more versatile/powerful than NAI because of Controlnet and all the LORAs available for it.If you just want quick, easy gens then NAI will give better results, but if you want to tweak and polish you can do way more with Autismmix and eventually tweak and polish where the results are better than NAI could give.So basically NAI for low-effort gens and Autismmix for high-effort gens.
>>101944972>You may not use the Output to train, fine-tune or distill a model that is competitive with the FLUX.1 [dev] Model.Wait does this mean? You can't use generated images to train with? How would they ever know?But you can finetune it? If something like pony flux was made, that's fine?
>>101946023workin on it, flux really likes to make women look like models, and it messes up the text when i speed it up with q4 instead of fp16
>>101946227Thx bruh
>>101946252>flux really likes to make women look like modelsOn the bright side, it could be worse: it could be a woke model that likes to make women fat pink-haired lesbians.So be happy it's not that.
>>101946277leave this general. chud. you're not welcome here.
>>101946252>flux really likes to make women look like modelsAre you sure?
>>101946252here, from first flux testrun when it dropped
>>101946308promplets absolutely unfathomably un fucking believably BTFO'd. you gotta catbox so you can teach them how you did it.hoping you don't get a vacation for this though.
>>101946252Would the average (non-jeet) person really fall for this? I can look at this and instantly tells it's AI but I've spent hundreds of hours looking at and genning 1girls
Going to train my first Flux LoRA, why are some of them ~20 MB and others ~ 2 GB? And how would I configure this in AI-Toolkit? Not too confident in a 20MB file to capture every detail.
>>101946236NAI has the best inpainter available currently. the only benefit pony has is lora training for characters. stylistically it's even worse because everything comes out looking like brown-tinted sludge. if you want actual eastern art, NAI is on top 10/10 of the time. if you want western-esque simpsons porn, then pony wins.
>>101946322flux got normie pics in data, just need to tickle it out.