Discussion of free and open source text-to-image modelsEmergency bake editionPrevious /ldg/ bread : >>101725030>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
finally
Flux is awesome
ty baker!
ack
I find the more I understand flux's prompting style and just what it can do the more I love it.Like the limitations put on me by the limits previous models are slowly being dissolved
>>101726656Flux is not finetunable. We have been over this many times.You will never see a massive finetune of flux and a flourishing finetune-mix ecosystem around flux. No NAI or Pony diffusion will ever come out.At the very best, some super small niche loras, but even that is at this very point impossible, as the tech does not exist and BFL will not be helping. They actually unironically think it's impossible too.What you see is what you will get. 10 years from now and flux will be at this same stage.
>>101727499teach me how to prompt flux effectively
I'm promplet, I will just use wildcards.
>>101727513Someone did a LoRA on 24GB https://github.com/bghira/SimpleTuner/pull/622#issuecomment-2267624531And there's this https://x.com/ostrisai/status/1820219528254488942
>>101727538>Someone did a LoRA on 24GBOh my.
>>101727514Just describe the scene in natural language
>>101727499Agreed. I think I still have a lot to learn, but I'm finding it much easier as time goes on.>>101727514Unironically ask an LLM to help you modify your prompt and learn from what it does. I guess because Flux is captioned by LLMs they seem to "speak the same language." It's not a silver bullet but it helps. I imagine some LLMs work better than others. I've been using whatever OAI slop Copilot uses these days. Most LLMs use OAI slop to make synthetic data so I imagine they all prompt pretty similarly anyway.
>>101727514think like a vllm
>>101727513Imagine being wrong. https://github.com/bghira/SimpleTuner?tab=readme-ov-file#flux1This is the same nigger who said it was impossible. Fuck off back to your den, Emad, don't you have 40 million dollars to launder this month?
>>101727594fp4 mayhaps?how shit would that be
>>101727594NTA but nice! I didn't think it would be possible for awhile longer.
>>101727584>>101727656The feels.
>>101727513no, it's finetuneable. just not easily on consumer hardware. if good loras can be trained on 24gb then the potential for a flux ecosystem is there. it needs to fit within the max consumer card really, as the next tier up is like $10k more.an actual finetune is another story and will be impossible except by big teams. which was always kind of the case anyway as you can count on one hand the amount of actual relevant finetunes we've ever had.
>>101727669Why can't we run flux/SD on multi GPUs anyway?
>>101727669Nobody is buying $10k GPUs to train LoRAs, they rent them for a few bucks.
>>101727594Well you will see. Distilled model will not respond to training the same way. There will be issues. It's not as simple as it is with some of the other models, some which supply training code and examples.Flux is not it and will never be it.Oh, and where is your negative prompts? Worry about that first, before you dream about massive finetunes with billions of images.
>>101727657
What prompt do they use with a vision model to describe images? Is it just as simple as "Please describe the following image." Or is it more complex? The closer we can get to how the original captions in the dataset were worded, the better.
>>101727714We may never know
>>101727681> Well you will see. Distilled model will not respond to training the same way. There will be issues. It's not as simple as it is with some of the other models, some which supply training code and examples.> Flux is not it and will never be it.Source?
>>101727714Copy an image into chat GPT and ask it to describe it and then prompt like would.
>>101727677probably because image models never needed more than like 10gb until now, so it was never a priority. there's a lot of scrambling after flux dropped without notice and BTFO all previous models in 90% of use-cases.
>>101727538Looks like schnell will be the most worked on.Can the dev profit from that or are they too different anyway?
>>101727729Me. You can quote me.
Why doesn't flux have a negative prompt? It's such a pain in the ass to see something undesirable over and over again and have to find some indirect way of discouraging it when you could just put one word in the negative prompt. I wonder if it is something that can be added later?
>>101727681>Oh, and where is your negative promptsWait, flux doesn't support negative prompts?
>>101727755Yeah, you know nothing. Distilled models vs trained models are exactly the same, it's just a difference how they are made.
>>101727681>here is evidence that it's trainable>you: picrelSeriously just shut the fuck up. You were wrong, deal with it. The community has massively taken to this model.
>>101727763the flux team said you don't need it and just need to prompt better instead
>>101727678I expect some price crash soon on the rent market too, once Nvidia sold to everyone and a bunch of overeager companies realize they actually don't need all these fancy A100/H100 and dump them in the market.
>>101727750It's the difference between SDXL and SDXL Turbo
>>101727750I don't see any noticeable difference between schnell and dev. People say schnell can't do text, but I was able to prompt for text.
how can I improve here?
>>101727787Flux pro has negative prompts.That statement of theirs alone should signal to you something about their mission.
>>101727814I have had limited phasing too. Woman inside the table
>>101727574That looks nothing like the prompt in the filename, the sad true is that all flux gens look the same
>>101727831It's understanding long prompts. The limit is still high in schnell (256 tokens) but 512 is pretty magic on dev
>>101727848They never said that though, he made it up
>>101727846Improve what?
>>101727869the workflow
>>101727846Be more verbose. Put your prompt into an LLM and ask it to make it more eloquent.
>>101727862>its not real because I choose not to believe ittrump voter?
>>101727856Yeah, I should cut Orientalism out because it really doesn't understand it. I'll acknowledge that even though I know you're trolling.
>>101727878>the workflowWhat are you trying to achieve?>A large, rotund gorilla sits at a computer desk in a thinking pose. The gorilla's massive frame fills the chair, with its hand resting thoughtfully on its chin. Its deep, intelligent eyes are focused on the computer screen. The gorilla's fur is dark and thick, contrasting with the modern, sleek design of the computer. Papers and books are scattered around the desk, suggesting a scene of intense contemplation. The background features a cozy, well-lit room with a window showing a lush jungle outside. The scene captures the unexpected blend of primal strength and intellectual curiosity.
>>101727875wtf is that
>>101727846There's a separate CLIP, seperate T5, CFG adjusting and neg prompt (not currently working) workflow. That should be the default that ships with ComfyUI, but for some reason he gives this extremely simplified one.
>>101727908Honey
>>101727912is there a picture on what its supposed to look like so I can attempt to reconstruct it?>>101727899>>101727883so basically the longer the prompt the better?
>>101727927>so basically the longer the prompt the better?the more descriptive the better
People don't quite grasp that this model is not like the research drops. This is model from a company with mission to make money, not enable basement coomers. These "open model" versions were not made so that you can finetune them properly.It's not in BFLs best interest that some other company could take their schnell model and finetune it to make it better than Pro and then sell access to that model. It would be disastrous.I trust BFL engineers when they say it can't be done. Not in the scale that people here hope. You will never have schnell perform better than current Pro.
>>101727935does it make a difference if you write like whole sentences compared to just words separated by commas that describe the scene?
I don't think this model is that good at generating pictures of women. Either that or I'm doing something wrong.
>>101727912All I have done to the default workflow so far is add an optional upscaler. Figuring out what the limitations of this model are first, and I am not hitting too many walls where I feel like I need to alter the workflow.
>>101727927https://files.catbox.moe/efb6nn.pngNote the dev has said lowering or increasing CFG 1-3 may help with following styles for longer prompts (default is 4), though that may just be the model's creativity going up a bit depending on how it interprets your prompt.
>>101727642>>101727476kek
Where I wish I was, but can't be because I am no longer as free as I was when I was 24.
>>101727952It uses both clip and t5 so you can do either. But it understands things like "A blue ghost wearing a red hat on the left and a black woman wearing a green dress on the right" which you can't do with the old way.
>>101727968I was pretty proud of that one
>>101728046kek
>>101727966>Note the dev has said lowering or increasing CFG 1-3 may help with following styles for longer prompts (default is 4)how to do that?>>101728010>It uses both clip and t5 so you can do either.can you do both?
>>101727956Flux has a default style for women, and it doesn't look very realistic. You have to play around with the prompt until you find something that works for you, also increase your steps, I always do 40, once you've figured out the sweet spot but it's easily the best model for genning women, just takes a while to find the right prompt though (depends what you're going for).
>>101728046hehe
>>101727956ai training on ai moment
>>101728065>can you do both?yes, but I just put the same prompt in both and use the t5 style
>>101728065>how to do that?Check the workflow I linked.
>>101728076I'm not sure if I even know what I want, but maybe something like this. It doesn't have an epic lighting, it's just a girl, you can see some noise, her skin imperfections, a scenery around her instead of a blurry background, the lighting isn't perfect. Any advice on that?
>imagination is the only limiting factor I'm fucked. It was nice knowing you.
Every day I get closer to buying a dedicated proompt machine
>>101728132>boring Snapchat photo circa 2015
>>101728136Where do you want to go? Gen that. For me, it is the mountains and the interior of a deep abandoned and long forgotten mine. Places I'll never go to.
>>101728145Yikes, if you have the money then just do it. I tried scraping by with a 1080 for like a year and it was torture.
>>101727444checked!>>101726868Does this even work? I don't understand funkopops.
>>101728152Another one I use "Snapchat photo of X, The photo is taken with bad lighting."
>>101728152Kek it works!>>101728145>>101728164People who gen with small GPUs or even just with the CPU are based. You are fighting with the tools you have. I was like that in 2022.
>>101728164I can do a 1024x1024 SDXL image roughly every 30 seconds, a Flux dev-fp8 1024x1024 image roughly every 6 minutes. I'm trying an anon's fancy ComfyUI Flux prompt right now which is taking extra long. My 2060 12GB was great for 768x576 SD 1.4 outputs like picrel, but showing its age now. At the same time, I want to hold out for cheaper 24GB cards, I feel like a 16GB upgrade won't future-proof me for long.
>>101726935This is what I get from:>The page appears authentic and ancient, and evokes a sense of indescribable dread.My love of books means I will be genning an outrageous # of books for really no reason.
>>101728226>6 minutes
>>101728233this book
>>101728190Thanks bro>>101728248What do you run and how much faster for you?
>>101728260How does it get fingers right?
are hand adetailers just gambling for a good one or is there a lora i can't find due to my own retardation
Does anyone here dare to run Flux on pure CPU? What's it like, an hour for a gen?
>>101728273You wouldn't have that problem with Flux.
>>101728284i also wouldn't have 1girl loli:1.5 bending over spread anus gaping anus style_anime either on flux
Thanks to the anon who posted catbox earlier, first time I've gotten ComfyUI to werk. Going to have a lot of fun messing with this one.
>>101728226>>101728248>1070 8GB>10 minutesBros... I don't want to work again... I want to stay a NEET...
>>101728266My 6950xt (all are 16gb) gets me just a little faster than you, as I have it configured, with fapper-eight.I may be able to try some things to get way better performance, though.
>>101728272By not being a coward like Stability AI and training on images of real people. SAI was so worried that people would use the model to generate images of children that they didn't release a single good model after SD 1.5 in 2022. They let Jewish financers and academics pressure them.
>>101728301Maybe the AI VC bubble will burst and a lot of start-ups will sell off A30's/A100's at bargain bin prices.>>101728310Good to know, thanks
>>101728314SD did train on real people. You can look into the laion dataset yourself.
>>101728076its because its trained on sdxl images, looks like the 1girl from juggernaut>>101727891i'm not trolling, once you have seen one flux gen, you've seen them all, look on reddit, facebook sd groups, whatever, all flux gens look the same specially the movie poster ones, also every gen posted has been cherry picked, (unlike sd3) so far it feels like another new ai generation company offering their new product, that in the end is just a game of playing gatcha, the less control you have the more slot machine feels likethats why i rate a comfy sd workflow over flux for now,
>>101728325Imagine if this scene was like that peeping cartoon where the peeper gets killed when the peeped turns out to be a horrific monster, haha
>>101728325It's excellent, it only makes some little mistakes, like it doesn't know where exactly doorknobs go.Here, it's excellent, but I never mentioned any picnic tables.
How do you see what's in the prompt of a png without getting the whole workflow when you import the image in comfy?
>>101728332Yes, that's why SD 1.5 was able to generate images of real women and remove clothes with inpainting. SD 2.1 and SDXL were heavily filtered and not nearly as exciting as 1.5. The community spent hundreds of thousands of finetuning in total only to get meh-tier models like pony.
>>101728344open it in notepad
>>101728314lol flux didnt train on real people, it used mostly artificial images, thats why all the women look like a mix of sdxl and dalle
>>101728359Oh it's in clear text! Thanks anonAny way to get it without having to change the \n to newlines myself?
>>101728260For some reason that pic reminds me of this
How does negative prompts work with natural text?You write a sentence or just tags?
>>101728376another anon.notepad++ iirc
>>101728393Guess I'll use that.
>>101728386Nice
I'm gonna catfish the shit out of dudes on tinder
burger king security camera footage
sorry.
>>101728410I like him.
>>101728416Quite hard when you can only generate one image per girl.
>>101728416Hard to catfish without working loras to produce a consistent face, soon.
Anyone else have this checkered pattern appear occasionally with Flux?
>>101728436>>101728435see >>101727538
>>101728438With low guidance.
>>101728406your tits, ma'amyour milkers, madamyour breasts, m'ladyyour mammaries, miss
>>101728438yes, its the safety watermark it pastes into your gen, or did you think you were getting all these images for free? mwahaha
>>101728448Get training then, I'll wait.
>>101728455For me it's her narrow waist, her pelvic lines, her cute blushing face, and her tits
>>101728448just because you trained an untested lora, that doesnt mean consistency anon
>>101728460how many sinks and mirrors does that bathroom need
>>101728390If you mean flux negative prompts don't work.
>>101728468>>101728461the same retard who said you can't train it is the one who added the training into his app. Jesus christ you SDJeets are pathetic.
>>101728469His, hers, and Jamal's.
>>101728474lol
>>101728471:^) we are having trouble getting a good Kamala.
>>101728454Ah yep, that did it. Was on 2.5 and getting it every few images, 3.5 is fine. >>101728460I don't care
>>101728502Go read what the fucker is writing. He is still very much unsure if it can work.He is not sure you can "improve" the model. He says it's possible to change the the model without completely ruining it. No one has yet demonstrated a true finetune and no one knows if it's possible. Not even the guy that made the finetuner.
>>101728539>No one has yet demonstrated a true finetune3 days anon...
Everyone screeching about Loras and fine-tuning. There's not a single solution of controlnet on the horizon. Controlnet is the only thing that has kept stable diffusion useful as a tool for so long. It's just a stock art generator until then.
>>101728547Yeah, the race is on, people with awesome gear are tinkering. I want one that focuses on buildings.
>>101728555Checked.Explain what controlnet is.
>>101728557I want one that focuses on big boobies and naked ladies.
>>101728555Also, no negative prompts.No negatives is a deal breaker for me.
Sometimes LLM just makes it worstOriginal prompt : Inside a kitchen, Trump is standing with a cutout of Hatsune Miku. Behind them, there is a window. From outside, Joe Biden is peeking inside through the window.Response : In this scene, an anthropomorphic Hatsune Miku cutout stands tall in a modern kitchen setting. The character exhibits a human-like posture and expression as it is positioned between Trump and the window. Outside, Joe Biden secretly observes through the glass pane, his curiosity piqued by the unusual sight within. The kitchen features sleek design elements, contrasting with the whimsical presence of Hatsune Miku's cutout. Model : Dolphin-Llama3:8b
>>101728565Actually, I would love one that could reliably avoid unsafe gens.How about adult nudity, no sex. That would be rather handy. ie art nudes.
>>101728539>IT'S IMPOSSIBLE TO TRAIN>OK YOU CAN TRAIN IT BUT IT'S VERY HARD>OK IT'S NOT THAT HARD BUT YOU CAN'T IMPROVE ITYOU ARE HERE>OK YOU CAN IMPROVE IT BUT IT CAN'T DO CONSISTENCY>OK IT CAN DO CONSISTENCY BUT IT CAN'T DO NUDITY>OK IT CAN DO NUDITY BUT IT CAN'T DO PORNSeriously you need to actually genuinely shut the fuck up.
>>101728569I hold out that someone will discover how to return the negatives (imo there are baked in negative prompts.
>>101728583Honestly creepy.
So I'm trying to understand rectified linear flow hereDoes the k-th rectified flow refer to a k-th training epoch?
>>101728590
>>101728582the more explicit a model is capable of the more artistic you can make your nudes. Just look at PonyThat said, Flux is pretty OK already at safe artistic nudes.
>>101728595No idea if that works, worth a try I guess.
>>101728661current century scrooge mcduck
>>101728650
>>101728685Broccoli shota
>>101728669this is nonsense
>dalle wont allow this prompthowever...
>>101728299
>>101728712pixel art prompt at the start makes some neat stuffpixel art, resident evil in game screenshot, miku hatsune is the main character, she has a black beret and gun.
>>101728564Controlnets let you guide the image in a very specific way. You can place line art of something in a controlnet and the AI will color it for example. You can put a specific style into the controlnet and it will replicate it. You can put a pose in a controlnet and the subject of that image will copy the pose. It's extremely useful.
>>101728706giwtwm
>>101728590kek, show me the image2video timeline too please, yeah it has been 3 days, so stop acting like is the 2nd coming of jesus, calm your panties anon, every new ai toy that has been released since in the last 3 years had this same hype moment and then people get tired of its limitationsthis anon knows>>101727936
>>101728731
>>101728590which ai model or company gan generate true consistency? thats the holy grail of ai generation and flux isn't either
>limited porn abilities I HATE THIS TIMELINE
>>101728742I don't get why you guys want to be peasants again
>>101728693Just tried and it kinda worked on my quick test.Prompt : Image of the skyNegative : Blue skyAnd I got consistent red skies.Without the negative blue skies again.
>>101728768I want nsfw of her.
Reminder that all the retards that said you can't train flux are nogens and vramlets who are mad they can't enjoy the model. There's literally no point engaging with them.
Could I create a LoRA with these images and then expect usable outputs?
>>101728770>it kinda workedpost the XY comparisons
>>101728788I can't train it but I'm still having fun>>101728789Probably. You'd want to split them into individual icons and blow them up to 512x512 first. I've made a Diablo 1 inventory LoRa with sprites only slightly larger that came out reasonably well.
>>101728777I tried.The prompt was her bending over with her heart-shaped ass facing the camera.Instead, I get smugly denied.
>>101728788>vrimlet
>>101728789I havent made loras with multiple instance prompts yet, but if you gave each one a unique name and images in individual folders, it should work fineive used kohya to make a character lora but that was just for one character. if each icon had an instance prompt and image associated with it, it should work?
>>101728801I want mostly suggestive too, but I'll wait for the fabled magical finetune I guess.
>>101728789I wouldn't do any training with less than 40 well-tagged examples. as far as the resolution though small is fine, no problems
>>101728801Try bathers renaissance painting
>>101728769it's just a meme, I wanted to make it jeb bush but flux didn't seem to know too well what jeb looks like
>>101728769US peasants? hAHAHAHAHAHAHAHAH PEASANT!
what the fuck is "variance explosion" and what does it imply for the result of the model?
>>101728798Yes, I wanted to split them anyway.>>101728819There are 20 in total.
>>101728791Not sure what you want, but it's easy to replicate.Picrel is literally with and without negative.It's pretty consistent for this at least.
>>101728789If will work if the model has a basic concept of the images. Otherwise, it'll either gen something that's vaguely similar but unusable or 100% likeness but completely unusable.
>>101728564>give model an image>it will arrange your gen along that imagedoesn't have to be an image, for example there are controlnets that set the camera position in your gen according to XYZ coordinates you give
>>101728849No way to know without trying; you can get away with 20 when training for a single person/face/object, might work with a single icon type as well.
>>101728868Wow you can control the camera?How long before you can do a proper video?
>>101728555Not a problem. Controlnets are easy to make and the concept is extremely simple. Let it progress one step at a time.
>>101728882You need a temporally consistent video model for that.
>>101728849>20between a rock and a hard placecooking too long gets overfit (gens adhere too much to examples), cooking too little gets underfit (too much variation, not consistently getting what you want)can you make more?(depending on use case, you might be able to commission a few more)
>>101728868You can make a controlnet for anything really, so long as you have an input image dataset and an expected output dataset.I've always wanted an Unwrapped UV controlnet that creates textures based on UV maps
>>101728890I prefer images anyway. So you can set a background, but adjust the camera angle, and then place figures and objects?
>>101728849>>101728897that being said, you can always give it a try. if you end up with something that works for your purpose then all's well. I'm just saying that I myself wouldn't
Requesting Mario saying NIGGER
>>101728940why are you so obsessed with them
female olympic boxing:
>>101728911I've only seen one controlnet of that type, there's another one on civitai to control light position, although it's spottynot sure why this haven't caught on, probably because popular frontends only have images as inputs for controlnetsyou can do all kinds of scene manipulation with CN, from placing objects to setting exact poses, face expressions, character proportions to making scenes from depthmaps to style transfer to full-fledged neural rendering (there are a few rigs for Mayba/Blender aimed at that)
>>101728940
>>101728802
What's the best way to get rid of jpeg artifacts in a gen?
>>101728731Yeah I'm really loving it
I'm starting to think when the prompt has nsfw context it goes spaghetti.> 'Slave attire, medieval bedroom setting, a single woman, displaying feet with detailed toes and soles, neat toenails, exquisite hands, five dexterous toes, lengthy brown tresses framing hazel eyes and smooth complexion, indicative of a 25-year-old, perky teardrop breasts, a petite yet athletic physique, bound by rope bondage while her arms are confined behind her back.'
>>101729028Is nintendo okay with this?
>>101729095Lack of training data, so yeah it goes to shit.
>>101728740giwtnwm
>>101729095>Anti-footfag measuresHoly chads
>>101729049click queue prompt and wait until a gen comes with no artifacts
>>101729051
>>101729049"jpeg artifacts" in negativesdependent on subject matter you might try "vector graphic" in prompt with a low weight
>>101729110Here's another spaghetti. I don't think the prompt is nsfw enough, but then I noticed "her legs spreading", this must be the forbidden keyword!> In this scene, a beautiful woman with Twintails, green eyes, and parted lips sits on the ground in a forest during daylight. She has white hair, long eyelashes, and wears a beautiful necklace. Her expression is serene, and she appears shy. From the front view, high skin detail, detailed skin texture, goosebumps, and a slight spread of her legs convey a sense of intimacy. The surroundings feature a gradient background, with birds, clouds, and butterflies adding to the fantasy theme. The painting style focuses on impasto, and the overall aesthetic is one of high quality and beautiful color details.
>>101728940Not condoning such language so here is Mario getting ready to rob some place with a brown Yoshi.
>>101729104>rectangle of coloured points resulting from a math equationnot their problem
>>101729124>the fand/hoot fan has entered chat
>>101729229
>>101729173
>>101729366
No need to wait over an hour for the next bread because it's already here...>>101729379>>101729379>>101729379
>>101729095rope bondage and the like just isnt in the data set, also "slave attire" is very vague and the model has probably not single picture tagged as such .. so ofc it goes haywire,also I am still sad FLUX doesnt know what shibari is
Running with an empty prompt in flux tends to give much more coherent results and and empty prompt with SD/SDXL. As in: the flux gens appear to be the product of much shorter, simpler prompts. What do you think this implies?
>>101729452>Running with an empty prompt in flux tends to give much more coherent resultthats cause dev/schnell are destilled models, they will go and converge to something, no matter what
>>101728214There's something about this that really gets me going.
>>101729280
>>101729419Try thishttps://huggingface.co/spaces/TencentARC/PhotoMaker-V2
>>101729489kek
>>101729681If the 'VIVE' said 'VIVEK' it would be perfect
>>101729681prompt?
>>101730095I too wish I had 50 caliber lever action M14s
