Discussion of free and open source text-to-image modelsPrevious /ldg/ bread: >>101836492>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
>mfw
>>101840360How did you describe the art style in your prompt? I like the texture of it
Blessed thread of frenship
Did the Forge cuck give us his tech already, bros? It goes so fast, bros. If he refuses to give it to us we should start a harassing campaign.
>>101840313I honestly don't think they scraped enough manga/artsy styles of anime from sites like pixiv, plus everything you ask Flux looks kind of synthetic. I think it's either a symptom of bad DPO or the model just doesn't know it.
What the. Interesting. Although it keeps getting the back wings wrong when she's not facing the camera kek. Maybe I'll stop genning with this character lora, I think I've proved to myself enough by now that it werks.
>>101840438it's already done, comfy-hating falseflaggereveryone's using nf4 in comfy nowhttps://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4
One thing I have noticed between fp8 and nf4 is my gpu runs 25 degrees hotter on nf4
>>101840473wild guess that it is doing less waiting for weights and more computing
>>101840448noice, dalle gen
>>101840438Can you repeat that?If you are asking if Flux works in Forge, then yes it does and he made it accessible for people with 6gb Vram etc to be way faster for them than before.
>>101840468Wow. This is so good, bros. Comfy is such a cool guy. Forge cucks get fucked! You have nothing now!
>>101840497you're so bad at this man
>>101840389anime pixel art, Amiga AGA, and PC-98but my cfg/guidance settings are taking the pixelization out unintentionally
I fucking hate people who use the word "bro" and "bros", ironically or not. It's so cringey.
>>101840530>amigathe comfiest computer of all time, niceI had one as a kid, no wonder I liked your image lol
>>101840448looks like it, yes, that first gen you posted looked pristine. at least we got some vroom.
>>101840490Dalle is clearly the best for character but is that Miura's style? I forgot to include that bit in the prompt, it was >Guts, berserk, closeup of bruised face, view from the side, by Kentaro Miura
>>101840537good morning sars
>>101840537bruv is crashing out
>>101840537ok bro
>>101840537wow almost as if he's doing it on purpose huh
>>101840494>with 6gb VramWTF. That's insane. And from what I've seen nf4 is closer in quality to fp16 than fp8. I'm tech illiterate, but with nf4 could we a 24B model that runs with 12GB of vram?
>>101840571>And from what I've seen nf4 is closer in quality to fp16 than fp8no
>>101840537bro is fine, but when someone says, "my guy", I immediately know that they're a faggot loser irl
>>101840592ok my guy
>>101840571yeah to me nf4 looks better than fp8.I don't know about the later part.
>>101840592What about "my brother in Christ"?
>>101840555>>101840556>>101840561>>101840562>>101840592Thanks for the (you)'s
>>101840592It's because they always use it in a condecending way, so you know you can't have any resonable discussion with the,
>>101840319>bad bay>no Miku slapping the cardMissed a chance to reference a dead meme.
>>101840619np bro
>>101840619no problem my guy brother in christ bro
>>101840592No it isn't fine. The only things that are fine are Anon, faggot, and retard.
>>101840629is this what family feels like?
Did anyone watch the Olympics Paris? I missed out, so I'm catching up using Flux.
I asked a LLM to design a mascot for /g/ and it called /g/ "4chan's tech-focused subreddit"
>>101840633you forgot your mother -- so fine, my bromigo
>>101840637I stopped watching when the blacks started winning everything.
>>101840625I tried but it kept doing this
>>101840637man that's insane physical coherence
Has any put together a list of characters Flux recognizes?
>>101840662They actually didn't do that great this time, even some of the sprinting races were won by white and asian menBlack people are slowly falling off, they're not cool anymore
>>1018406741. Donald Trump2. Hatsune Miku
>>101840674It "recognizes" a lot but can't reproduce them accurately. Like Rey from Star Wars, it gets her hair and outfit vaguely right but face wise looks nothing like her
>>101840674It recognises a lot of people but not trained enough, for example if you type Angelina Jolie you can tell it's trying to make her.But yeah we need a list of ones it does well.
>>101840674>>101840687Sonic, Mario, Meghan Markle, Queen Elizabeth, Hillary Clinton, Joe Biden, Vladimir Putin,
>>101840667Tbh the fact that she says "slaps" kinda makes me imagine that she is literally saying her stage directions word-for-word while doing them.
>>101840571nf4 is just as good imo, my only current pain is experimenting with prompts, the fp16 t5 doubles the gen time on my 3070
>>101840552kinda gets it
>>101840667>"so much fucking"What did she mean by this?
>>101840571>but with nf4 could we a 24B model that runs with 12GB of vram?It should be scalable, yes, and in the future we will likely get 1.58-bit models as well, which should also be much faster, because they will work only with additions instead of multiplications. It's no outrageous to assume in a couple years we could be running 70B diffusers with consumer hardware. That being said, I think we will start getting diminished returns very soon. Flux can get better, but not much better, and I think the bottleneck right now is in the dataset anyway.
>>101840738you already know
I just want to upscale images, what do I download
>>101840733Fits well, probably could have had him as a friendly baker instead of a science teacher and the story would have worked well.
>>101840771Forge will be the easiest thing, but comfyui will have less overhead.
>>101840712kim jong un
What the fuck is nf4 now? Is it a single file like the fp8 version or do I need additional files like fp16. Yes I'm a retard and don't know how any of this works
>>101840771https://github.com/upscayl/upscaylUpscayl is normie proof
>>101840738Hatsune "so much fucking" Miku is this one. You need a big GPU for her.
>>101840737Not bad. Would be nice if Hunyuan devs released a 10B model.
>>101840831She's a big girl
>>101840771Download a model from here to fit your purpose:https://openmodeldb.info/And use one of the programs listed here:https://openmodeldb.info/docs/faq#maintained-programsThe program and the model are all you'll need to get started
>>101840811I'm making shit up so don't believe anything I say, but I believe it all comes in one file.
>>101840282Well, I tried to get it to draw correct nipples but got grossed out after too many illustrations came out with eyes, mouths, blisters, puncture wounds, incisions, and who knows what else in place of nipples. Every time you neg'd one gross thing it just substitutes another. Excellent work, faggots.
>>101840811single file, fits into 12gb vram, works on as low as 6gbvram, works in both comfy (with some fuckery) and forge (default)
>>101840869>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/aco/sdg>>>/trash/sdg
>>101840892/trash/ my beloved
>>101840831She at all the datasets
how's SUPIR?
>>101840890>with some fuckeryWhat you mean some fuckery, bro? I thought Comfy had our backs.
Blind test. A or B anon?https://imgsli.com/Mjg2NTc3
>>101840869Also the quality of the illustration was very poor. Sub-SDXL. Like it had literally never seen a pen and ink drawing in its life.>>101840892Deflect all you want, retard. People like looking at naked people.
>>101840811If you're using Comfy, you will get better results using separate t5xxl fp16, the current nf4 model has a limpdicky useless t5xxl fp8 baked in.
What is the current state of image to video in Comfy? I remember a few months ago some very cool things were coming out, but I've never seen anyone posting videos here.
>>101840394very good
>>101840929I like B more and I also suspect it adheres better to the prompt.
>>101840952All I know is that the creators of Flux have a text to video (I assume image to video will work with it too) on the way for local users.
>>101840937Are you using a specific prompt that gets you the limited colour palette without being totally greyscale?
>>101840981How. I saw your post yesterday too and it looks pretty coherent.
>>101840976Thanks.
That's a nice t-shirt
>>101840786>>101840817>>101840849Thanks
>>101840977They both adhere accurately to the prompt. People keep sayng nf4 is better but I don't really see it, so want to know peoples opinions without bias.
>>101840981cool looks 3D when I cross my eyes. Now my eyes hurt.
>>101840996I like how the "mom" look slike a tranny too.
>>101840831
>>101841018And the dad identifies as an Alien. (that's a mask)
>>101841001If it was me, I'd say that he just got a tatoo of a shirt on him then wears a piece of string that looks like the shirt top around his shoulders.
Comfy nf4 node was updated to "Fix memory issues."I am still waiting for loras to work tho
>>101841030That's obviously an alien, anon. The dad is on the far left.
>>101841015Zoom out or move away from your monitor, it's easier on your eyes.>>101840995I use IW3 to convert images and videos to 3D:https://github.com/nagadomi/nunif
>>101841050good idea
Wasn't forge telling everyone to go back to A1111? What changed?
>>101840992It's finicky and not reliable, but how this happens is I put "Drawn in pencil by Tsutomu Nihei." in the clip prompt, and the rest of my prompt in the t5 prompt. Some gens turn out with no color. Most don't even have Miku.
>>101841063Flux inspiration
>>101841035I updated and it still juggles shit in memory for a loooong time between gens, I guess it will take a few days and updates to become actually useful. I'm not sure the overall time is shorter than with fp8 right now.
so ponyflux when and where?
>>101841063
>>101841104Dev has cucked license and Schnell is shit. Not happening.
They finally caught the villain.
>>101840712>Hillary ClintonI wish but no.
>>1018411042 weeks>>101841131Wrong, you didn't read the license.
>>101841127Is that flux, looks like it know Brosnan pretty well.
>>101840929A, the facial expressions are more expressive, the colors/shading is more natural, and the background is better
>>101841139old man withers looks like THAT?!
>>101841141
>>101840933Oh and while I'm here, the model output its core red-shifted censorship heat map fwiw. You're gonna need to eventually fix how the model interprets prompts related to all these areas, if that's what you're into. Anyways, good luck.https://files.catbox.moe/pddnfo.jpg
Now this is a raw epic 4chan mascot.
>>101841159i had to scum for it quite a bit, my other gens were not so accurate
>>101841165The more you know
>>101841173belly button censorship baka my head
>>101841173What did you prompt for that style? I like it
>>101841177Lol that actually looks like Pierce Brosnan and John Cena combined.
>>101841202Damn, Flux is so good at making sexies.
I bet Flux can't generate a cute Italian gymnast doing the splits on top of wheels of Parmigiano cheese.
>>101841193Because the straight line that forms the navel runs parallel to the pussy so it will prevent the formation of certain shapes in that general vicinity so the model doesn't draw a gash by accident (or on purpose) just by echoing or shifting around shapes in the adjacent region basically.
>>101841255you're in too deep, anon
>>101841255you have some pretty schizo theories of how censorship is done in flux
>>101841240I can try, but she has to be 18 plus ok?
>>101841255I believe you
>>101841296Forget it.
>>101841306
>>101841316>when she nut and I'm are succing
gibs me Flux pro free SaaS website now. or do i have to have enterprise gpus to sample any new open source stuff that drops?
>>101841225
>>101841365I want to go BRRRRR on that ass.
>>101841373>>101841365>>101840892
Porn strikes fear into the zoomer.
This is a Fox News alert
>>101841365So soft, yet so firm, so smooth yet so much to feel. I can use her ass as a pillow and never worry about her farting because she's an android.
>>101841410I want to see Flux generate some hangers, true zangers. Not this bolted on shit.
>>101841410I want this is full video animation one day, I will make it like those JAV movies where they have the anchor woman continue reporting the news as men keep cumming on her face one by one.
>>101841442SHE BROUGHT THE CHEESE!
>>101841442boneitis at such a young age, so sad
>>101841457Maria Ozawa...my first love
>>101841435>>101841373>>101841392
>>101841457imagine what we will create once we can AI generate convincing real-life pornsociety will grind to a halt as all men become 24/7 gooners
https://civitai.com/images/23847350It's happening
>>101841392I caught the Villain, I'll send her to over there.
Is there a way to improve Flux-dev performance on a 4080 Super? I'm using flux1-dev-fp8 and I'm getting 27 seconds per iteration.
>>101841494>no couch
>>101841173If its true, we will begin to see cracks but not on flux's censorship shield.
>>101841476Ah I wasn't sure if we had a man of culture here, but you have my respect.Also I want to recreate such movies like MVSD-248 Cum Buffet
>>101841457>>101841491>>101841525>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/aco/sdg>>>/trash/sdg
>>101841506Post your workflow, I have a 4080 standard and get 1s/it
>>101841476Ah I wasn't sure if we had a man of culture here, but you have my respect.Also I want to recreate such movies like MVSD-248 Cum Buffet>>101841491The future is scary because I will have to gain more self control when this is possible....especially when it comes to VR simulated worlds...
>>101841533Excuse me, but you missed this poster as well.>>101841493Kind regards,Anon.
>>101841536dunno what that is but i want it
>>101841547Thanks I've been working real hard to mod this place, I wish I got paid, but yeah I somehow missed that and I feel great shame...
>>101841152Yes I did. You can't make dev finetune and sell it as a service, which is exactly what any big finetuner would do, like NAI or the ponyfag.
>>101841506Brand new https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4 and corresponding checkpoint might help.>speed improvements ranging from 1.3x to 4x compared to FP8, depending on the GPU and software setupbut it's not extensively tested by people yet
>>101841548I still woul- no... I will.
>>101841573Wow lol. Also I like the reflections.
>>101840571>And from what I've seen nf4 is closer in quality to fp16 than fp8.not truehttps://reddit.com/r/StableDiffusion/comments/1eplvi8/the_image_quality_of_fp8_is_closer_to_fp16_than/
>>101841548As an ASS-MAN, this is the perfect woman.
>>101841442that's impressive, Flux is so good at anatomy and poses, finetuning this shit for porn will be a blast!
>>101841595>one image of a trainWe've seen many more examples here where nf4 > fp8.
>>101841595Yeah, but everyone in the thread pretty much agrees that NF4 looks better overall - so 'accuracy' towards FP16 really means nothing if it enhances the image.
>too dumb to remember the speed & ram usage of previous version
>>101841571btw pic unrelated, it's just citiviai being broken (the checkpoint was supposed to unlock with donation goals, instead it now costs *more* buzz)
>>101841617>maybe more
>>101841595Sadly OP only uses one image as an example. So maybe some images are closer and some are not depending on the seed and prompt.Also supposedly it's closer in Comfy and only seems more differeny in Forge. hmmm
>>101841617Why are you using SD3?
>>101841548women have reached their final form, where their pussy and ass merge into 1.
When creating photorealistic stuff I can't seem to stop Flux from creating a sharp focal point while everything else in the image is a blurry mess even when I put "sharp focus" in the prompt. Any tips?
>>101841625that's not how it works, the only question you should ask yourself when comparing quants is "is this quant closer to fp16 to the other quant?", that's only what matter, how much accuracy you lost compared to fp16 because fp16 is the thing you should aim at
>>101841617Now I am become goon, the coomer of cooms.>>101841628In the examples I saw here, nf4 is not just aesthetically better but adheres more to fp16 results than fp8
>>101841595vramlet copers don't care about the truth.
>>101841628>everyone in the thread pretty much agrees that NF4 looks better overall - so 'accuracy' towards FP16 really means nothing if it enhances the image.That's fair, it's funny to say that making flux lose accuracy helps him to make it more "beautiful" pictures though
>>101841648It's not that simple though, smaller is not automatically worse. Nf4 is quantized far more intelligently than fp8 which just mindlessly throws away 8 bits on everything.
>>101841674very impressive result if this is just a pure flux gen
bitsandbyte slowing down my overall gen time, it's over
>>101841653>nf4 is not just aesthetically better but adheres more to fp16 results than fp8this is not true at all why are you all repeating that lie? we also have eyes and can see through your bullshit
>>101841646Dynamic thresholding+CFG to improve prompt adherence, plus stuff like bokeh in the negatives helps a lot for me. Realism lora is good too.
>>101841666Maybe someone should try perturbation on flux model and check if it affects it in any way.https://www.reddit.com/r/StableDiffusion/comments/1dfuicw/perturbed_sd3_experiment/
>>101841687
>>101841689>posts the train image againsee >>101836843>>101837061
huen/beta seems to be the most optimal combo from my anecdotal experience, but it's so fucking slow
>>101841641https://files.catbox.moe/kj25yb.png
>>101841675>Nf4 is quantized far more intelligently than fp8 which just mindlessly throws away 8 bits on everything.it is true, but the reality is still that fp8 gives closer outputs to fp16 than nf4, "dumb 8bit quant" beats "smartly arranged 4bit quant", and to me that's not surprising, on LLMs everyone agree that nf4 gives worse results than fp8
>>101841723why have you got guidance on your positive but not your negative
>>101841689Come to think of it, is your train image using the forge fp8 text encoder vs fp16 encoder in the other images? That would explain a lot.
>>101841720>still missing the pointI'm not arguing whether nf4 produces better pictures subjectively, you have the right to believe that nf4 does and I can't say anything about that, art is subjective. What I'm arguing at is about the cold and objective accuracy comparaison, does nf4 give pictures that look like more like fp16 than fp8, the answer is NO
>>101841751>you just know
>>101841723Fix your shit. I refuse to believe that flux has that anatomy.
>>101841698Thanks anon.
>>101841755Because negative doesn't matter, it's un-used in flux, but the ksampler needs an input for it, and the clip node needs an input for it.
>>101841796Oh right, you aren't using any of the fixes.
GUYS I somehow got a NUDE image with flux!!https://files.catbox.moe/odgc7x.png
>>101841777nf4 fp8
>>101841796>Because negative doesn't matter, it's un-used in flux,we can use negative now anonhttps://reddit.com/r/StableDiffusion/comments/1emy7uv/negative_prompts_really_work_on_flux/
>>101841824ayy lmao
>>101841540https://litter.catbox.moe/du2uqv.json
>>101841760>does nf4 give pictures that look like more like fp16 than fp8, the answer is NOthe answer is YES
>>101841828I am aware but I really can't be fucked with halving my s/it times for a negative prompt I don't need nor will use.
>>101841824top kek
>>101841072>Drawn in pencil by Tsutomu Nihei.Thank you.
>>101841824>>101841843/x/ bros???
>>101841843can you provide the workflow for the 3 of them? that looks interesting
>>101841827nf4 fp16
>>101841827Atrocious at dynamic poses. Dall-E 3 crushes it at those, though hard to compare because of the shit filters it has.Try to do a fully clothed male. Maybe the schizo posters are right and there is some filter layer that activates and fucks up the image on purpose when it detects "nsfw".
>>101841837Change this from default to an fp8 weight mode.
>>101841847so don't say that "negative doesn't matter and is used in flux" then, it matter and can be used, it's just that you don't want to deal with the speed drawback, and it's fair no one want slow shit, you can still improve that speed a bit thoughhttps://reddit.com/r/StableDiffusion/comments/1enxcek/improve_the_inference_speed_by_25_at_cfg_1_for/
>>101841837>>101841890Also I just realised you have the fp8 safetensor, use the "load checkpoint" node instead.
>>101841861nta but here are my catboxeshttps://files.catbox.moe/2jtb3b.pnghttps://files.catbox.moe/t0rkrr.pnghttps://files.catbox.moe/5gez8v.png
>>101841891It doesn't matter. I will say it again, negative does not matter in flux. I don't need to take a speed hit to add "black skin" to the negative when I can just add "white skin" to the positive. You are all coping and stuck in the past. Negative doesn't matter in flux.
>>101841908Cool.
>>101841908Damn nice
Itadakimasu
comfy flux dev at fp8 weights + fp16 text encoder, or forge np4? i'd assume comfy would have better text encoding but is that the case? i've had good results with np4/forge.
>>101841965*nf4
>>101841943Then you just haven't encountered a situation where you needed them yet. I've had plenty of gens where flux sorta followed the positive prompt but not entirely, and I had to use a negative to push it further into doing what I wanted.
>>101841943>Negative doesn't matter in flux.it does matter, don't make your delusions a reality to the others>>101841943>You are all coping and stuck in the past.You're the one stuck in the past, using flux at CFG = 1 is retarded as fuck because you get way better results if you increase that shit and play around with the negative guidancehttps://imgsli.com/Mjg2Mzgy
>>101841965if you have enough vram go for fp8
>>101841890Whoa, this is amazing! Thank you so much!
i'm trying to follow this workflow but i'm using nf4. can i still use these clip files?
>>101841646I'd recommend this instead >>101736456
>>101841991>You're the one stuck in the past, using flux at CFG = 1 is retarded as fuckI have played with Dynamic Thresholding and gotten good results. But I still don't give a shit because it's not worth the performance hit.>it does matter, don't make your delusions a reality to the othersIronic.
>>101841991>https://imgsli.com/Mjg2MzgyThis is just dynamic thresholding and has nothing to do with what you put in the negative prompt.
>>101842039yeah, but you need the new checkpoint loader for nf4
>>101842063>has nothing to do with what you put in the negative prompt.if you want some negative prompt examples I have some
>>101842063>This is just dynamic thresholding and has nothing to do with what you put in the negative prompt.but his point was to say that getting that speed lost because of CFG > 1 isn't worth it at all, it is actually worth if you want to use negative prompt, have better adherance overall, and also better style adherance thanks to NegativeScale at 10
>>101842098OK, using https://imgsli.com/Mjg2Mzgy as an example.You have negative cfg at 10. If you put Hatsune Miku in the negative prompt, she should disappear completely. Do this, and show me a result that is't explained by random noise. Also, you can simply put white skin in the positive for your trump pic.
>>101841824WOULD
>>101842118>If you put Hatsune Miku in the negative prompt, she should disappear completely.holy retardation, who would put the same prompt in the positive and the negative prompt?
>>101842128You have it weighted significantly differently, we should see a result. Show me.
>>101842118>implying negative prompt should somehow take completely precedence over positive promptretard
>>101842118>>101842128Yeah, I don't think you could even do that under normal circumstances and not see significant leak of the character in the negative prompt.
>>101841824is this AI?
>>101842141why would you want to do something like that in the first place? it has zero real life use case, no one ever puts the same word/concept on both positive and negative prompt, are you retarded or something?
>>101842146Yes it's AI(It's actually not, Aliens are real...)
When did this become /sdg/?
>>101842165always was
>>101841944>>101841945Shame about the compression effect, I keep getting that randomly for some reason even at 35 steps.Using SD1 just destroys all detail, 4 channel VAE and all: https://imgsli.com/Mjg2NTg1
>>101842165When you stopped posting images.
>>101842128>>101842144How about this then. Show me a single use case for negative prompting that simply cannot be achieved in the positive.
>>101842181I don't have a need to shit the thread with billion images of the same thing. I post only when I have something cool to show and that does not happen every 5 minutes.
>>101842184Why don't you do the same? You're effectively saying that negative prompting is always useless regardless of model, because you can enforce things in the positive.
>>101842200Skill issue.
>>101842184CFG > 1 (aka speed decrease) is interesting for 3 reasons:- Better prompt adherance- Negative prompt- Better style adherance if you tinker with GuidanceNegyou have the right to believe that negative prompt is useless, but CFG > 1 gives you more possibility than just that, so it's not that simple
Bon appetit
>>101842236this too, the built-in guidance is like a bandaid that's not nearly as effective as true CFG, setting CFG>1 forces it to try harder to follow your prompts
what is t5xxl? additional details/instructions?
>>101841573Its funny. In real life, his left hand has problems but in the mirror world it is his right hand(from the reality pov).
>>101841937Cute!
>>101841476>>101841543Fellow JAV enjoyers?
>>101842347flux works with 2 text encoders:- clip_l -> pretty good at tags but not natural sentences- t5xxl -> really good at natural sentencesThe goal there was to get the best of both words by combining the two
>>101842365Nice seems like Hunyuan makes the best asians by default.
>>101842347I believe clip l takes standard tags as input and t5xxl takes boomer prompting, but so far it seems fine to just give both a boomer prompt
>>101841617>>101841827>>101841871Are those positions not allowed?
>>101842377Does it beat Kolors?
>>101842373>>101842380thanks for the insight i was fumbling by treating both as flux's new style of prompting.
>>101842406No idea.
>>101842215> You're effectively saying that negative prompting is always useless regardless of modelNo I'm not, I am saying there's no point to it in Flux. >Why don't you do the same?ok, now you
>>101842450>still humans in the background of his image
>>101842450Shitty promptlet
>>101842450the "there at no XXXX" doesn't always work, that's why adding a "humans" on the negative prompt will make that effect bigger and really force flux to not draw humans, don't see negative prompt as something that compete against positive prompt, it's a teamwork, when one can't do it, the other can help, better choices and bigger chances of succeding with more help is always welcomed
>>101842461>>101842479Show me your negative prompt removing them entirely, I will wait.
https://files.catbox.moe/opw1w9.jpgI'm tinkering with the max_shift from ModelSamplingFlux (that shit looks like it's changing the color saturation and contrast), which value do you feel is the sweet spot (default value is 1.15)
>>101842450What the hell is this example. I asked you to provide a conflicting positive prompt. For example 'A photo of a populated city street full of people. There are no humans.' would be the equivalent of what you're asking others to do - which is to put a negative prompt that conflicts with the positive prompt and see what comes out.>OK, using https://imgsli.com/Mjg2Mzgy as an example.>You have negative cfg at 10. If you put Hatsune Miku in the negative prompt, she should disappear completely.You just prompted twice, once for a populated street and once for an empty street.Are you retarded?
is flux heavily censored or something? i can't get
>>101842518And I asked you to show me something you can only do in negatives. You have provided nothing so I accept your concession.
>>101842450You can't say "busy street" and "no humans" on the same sentense, you just conflict two concepts, and confuse the model.>I want the color green but redthat's not how that work
>>101842377It does have best photoreal selfies but you have to know how to prompt it and it only works with 768x1280 res.
>>101841843Milusex>>101841858The Demigure awakens!
>>101842533> >Why don't you do the same?>ok, now youSo you failed to provide a suitable example of your ridiculous usecase, which was a stupid strawman anyway.I accept your concession, it's pretty clear you lost here.
>>101841965>free speech and anime for allFUCKING FINALLY!
>>101842548Kawaii desu ne.
>>101842550>here's an example>no I don't accept that example therefore I win
>>101842535
>>101842566The example that doesn't even do what you said you should do?Yeah, I wonder why people would have a hard time accepting that.
>>101842548Hunyuan 1.0, or 1.1, that is.
>>101842583It's using the positive prompt to remove elements, without the use or need of the negative prompt, which circles back to my original point that made you become so obsessed, the negative prompt in flux does not matter.
Anons...I finally kissed a girl.
>>101842596It doesn't use the positive prompt to remove elements, because your positive prompt didn't establish the elements you removed in the first place.>Prompt 1:>A photograph of a bustling city street.>Prompt2:>A photograph of an empty city street. There are no humans>i prompted for an empty street and i got an empty street, therefore negative prompting is uselessare you daft?
>>101842566your example doesn't work, you ask the model "busy street" and a busy street has a lot of humans in it, and then you say "without humans", that's fucking retarded, you're literally saying.>Hey, I want a busy street so it means there's a lot of cars and humans, but I also don't want humans, sounds simple enough, right?No it's fucking not, and everyone would look at you funny if you were to say something like this in real lifeGo for some concept where you can remove something out of it and it's possible, like "A human without hair", flux will assume the human will have hair at 95% of the time but you can say "without hair" because Flux knows that humans can't have hair so you're not confusing it, your example is retarded because a busy street cannot have no humans in it, do you get what I mean by that or?
>>101842616Literally show me 1 (one) example that you can only achieve with the negative prompt. Prove to me that the negative prompt in flux is necessary.
>>101842450>No I'm not, I am saying there's no point to it in Flux.You're missing the point. It's more nuanced than that. Naming something can affect how effectively you can exclude it.>>101842596The negative prompt gives the user more power. Even *if* every concept had an exact anti-concept that could be put into a positive prompt (not the case), there are still the side-effects of the added connotations of those anti-concepts.>>101842631>Prove to me that the negative prompt in flux is necessary.>Prompt: Create anything except don't create a green zebra>inb4 skill issue
>>101842631No, you provide an example that actually works first. Prove your retarded theory correct.>OK, using https://imgsli.com/Mjg2Mzgy as an example.>You have negative cfg at 10. If you put Hatsune Miku in the negative prompt, she should disappear completely.As I said, you first. Go prompt>A drawing of Hatsune Miku skateboarding in New York, night, neons, 50's comic book style, there is no Hatsune Miku in this image.and get back to me. tell me how that works out.
>>101842649>Create anything except don't create a green zebraCool, now use a negative prompt to impact this.
>>101842653Simple>A drawing of skateboarding in New York, night, neons, 50's comic book style
>>101842669>Positive Prompt: Create Anything>Negative Prompt: Green Zebra
flux can do basically any scenario, it is open source dalle essentially.BUT sdxl/pony still do anime characters best as flux doesn't have them all in the dataset or the all the loras (yet)
>>101842677That's genuinely not the same thing at all.>You have negative cfg at 10. If you put Hatsune Miku in the negative prompt, she should disappear completely.Again, the equivalent of this for the positive prompt would be>A drawing of Hatsune Miku skateboarding in New York, night, neons, 50's comic book style, there is no Hatsune Miku in this image.Go do that and get back to me. Or admit you made a stupid fucking example.
>>101842631>Literally show me 1 (one) example that you can only achieve with the negative prompt. Prove to me that the negative prompt in flux is necessary.There you go
>>101842696Relevant meme
>>101842698My original point, that started this whole retarded argument, is that the negative prompt in flux simply does not matter. I can achieve anything in the positive. If I want a picture of hatsune miku skateboarding with no hatsune miku, I can simply take out the word hatsune mikuYou CAN NOT add a negative prompt of hatsune miku to remove her from the promptThat's the whole point.
>>101842721>ohoho my original point my original pointso you admit your example was retarded then?thanksretard
>>101842721>the negative prompt in flux simply does not matter.-> >>101842699
>>101842699
so nf4 in forge is more efficient in terms of memory/vram, but flux dev + fp16 text encoder is best quality overall in comfyUI, right?
>>101842699this.it is settled.
>>101842739this is a bed retard it just has pillows over it
>>101842757Is that how you live?
>>101842739you are incapable of abstract thoughtyou are incapable of admitting your ideas lack nuance
>>101842699You changed the positive prompt. Prompt for a bedroom and put bed in the negative.
Is nobody baking? I will if not
>>101842785he can't accept that this is the only fair comparison.hence why he cried and cried until he gave up>>101842677 >>101842721 >>101842732he can't accept he's wrong. can't accept he's a retard
>>101842785Both mean "a room without a bed" and the one with the negative prompt is the only one that actually did what you tried to convey, but yeah I'll go for bedroom if you want, just wait a sec
positive: mannegative: penis
there, now he's the olympics guy
>>101842821POV you just talked shit about Israel
>>101842804You are all missing the entire point.You are thinking about negatives. A room without a bed can be described simply as a room. It doesn't need the negative.
>>101842809Your image is not of a bedroom with no bed though. It's just a sitting room, which can be achieved by prompting "sitting room" in the positive.
>>101842836okay now prompt bathroom without a sink retardextremely obvious negative prompt options that you're so desperate to ignorego on, prompt 'urinal' or some shit you fucking bastard
>>101842843not just a sitting room doesn't have a bed, that's the thing, you restrict flux's creativity if you force him to go for only one specific room even though he could have the choice of many different rooms if you simply specify "bed" on the negative
>>101842836>>101842852How do I know you two aren't the same person?
>>101841937Prompt and model?
>>101842860>>101842860>>101842860>>101842860>>101842860
>>101841937Perfect for a childrens book.
>>101842867>gets proposed a reasonable example where a negative prompt could be applied>uhhh how do i know you aren't just samefagging!!!11!just admit you've got nothing.
>>101842696Not just anime. Flux has basically removed complete control for artists and fine art. And its base seems to have been tuned on AI slop rather than real artists.
>>101842836>A room without a bed can be described simply as a roomI'm guessing you can't keep a job very long, can you?
>>101842852>okay now prompt bathroom without a sink retard
>>101842883He's not me, I am literally the only person arguing that negatives aren't necessary.
>>101842870flux dev, "Art by Beatrix Potter"
>>101842942Just flux, with no extra model or anything?That's pretty incredible
>>101842910this is a retarded way of prompting, what if there's a room with 100 objects in it and you just want to remove one, will you go for that retarded prompting?>I want a room with this and that and that and that and that... 99 timesOr will you go for something more elegant?>I want a room>Negative prompt: "that only thing"
>>101842910NTA but I said this ages ago and you never addressed it properly. >>101842450This strategy isn't any different than on any other model. Flux isn't special when your alternative to negatives is literally just "don't mention the things you don't want".
>>101842958yeah raw, all the available loras so far are garbage and not worth using
>>101842964Still does not prove that negatives are necessary, there is not a single example you can provide that I cannot prompt using only the positive. Thus, my statement, that I made offhandedly to an anon who asked why my negative wasn't connected to a cfg, the negative prompt does not matter in flux. Particularly because I am using cfg of 1 so it does literally fucking nothing
>>101842969>Flux isn't special when your alternative to negatives is literally just "don't mention the things you don't want".doing "I want this without this" is less effective than using a negative prompt, it's just how it is >>101842699
>>101842988HuhSeems I have to try this flux thing.Thanks for the info!
>>101842993>Still does not prove that negatives are necessary, there is not a single example you can provide that I cannot prompt using only the positive.keep ignoring that picture, you're doing great >>101842699
>>101842884Its going to take time for people to let this sink in. It only happens after hitting many walls.
>>101843011I literally posted a room without a bed. two of them.
>>101843017the first one has a bed >>101842757the second one isn't what I asked, I asked for "a room" not a "sitting room", that's absolutely not the same thing, it's as if I would ask you to give me some pop suggestions and you will say "There's only Michael Jackson and nothing else"
>>101842884Yeah I prompted for an oil painting of a man wearing pajamas floating in space and it gave me this. It sucks at mediums.
>>101843031And I said there is literally nothing you can achieve with use of the negative prompt that cannot be achieved using only the positive.Show me a single example of something that the negative prompt is mandatory for and I will concede. I am not talking about something being more convenient to do in the negative, I am talking about something that is literally impossible without a negative prompt.
>>101843057you already lost. go back
>>101840929A. Which was which?
>>101843057>I am not talking about something being more convenient to do in the negative, I am talking about something that is literally impossible without a negative prompt.for the 4time, this picture >>101842699Without negative, you are forced to specify one specific room without bed even though there's more possibilities than just the sitting room, your answer is incomplete without the negative prompt, this is a fact, if you feel that's not important enough, you do you, for me it's important to let flux the freedom to choose the room it wants to draw, I don't want a sitting room, I want any kind of rooms, and only negative can achieve that, comprendo?
>>101843077B was nf4, A was FP8
hell yeah
>>101843091Literally just prompt for a room. What the fuck man. You also changed the positive prompt, that is not something that could have ONLY been done with the negative prompt.
>>101843118>You also changed the positive promptare you a retard? the negative prompt exists so that you won't have to write "without a X", why do you keep missing the point, "a room without a bed" and "Pos: a room + Neg: bed" LITERALLY MEAN THE SAME THING YOU FUCKING MORRON
>>101843134Are you retarded? You are arguing something I never said or believe.MY ONLY FUCKING STATEMENT IS:The negative prompt in Flux is not necessary because you can achieve everything in the positive"You DO NOT have a single example that I cannot do in the positive.Jesus Christ"You lose because of this different thing ha checkmate"Absolute donkey.
>>101842942Thanks
>>101843175>You DO NOT have a single example that I cannot do in the positive.I'm not going to repeat myself, you're wasting my time with your retardation -> >>101843091
>>101843189I am going to repeat myself, a fucking sitting room
>>101843175This is really silly. Again, you haven't addressed why, if this is such a revelation - that we use negative prompting for every other model.>>101842969Are you suggesting that everyone else is doing it wrong with negative prompting? We should simply not be using it since you can achieve everything in the positive?
>>101843189
>>101843178radevery image model I've ever used knows Potter, so it's an easy test
>>101843209I DIDNT ASK FOR A SITTING ROOM I ASKED FOR ANY KIND OF ROOM THAT DOESNT HAVE A BED IN IT WHY DO YOU KEEP MISSING THE POINT KILL YOURSELF NIGGER YOU ARE TOO RETARDED TO LIVE
>>101843212Because every other model is stupid, it doesn't use t5, it uses tags exclusively.
>>101841734sovl
>>101843226okay, just don't tag the things you don't want. basically what you've suggested.Are you telling me that if you prompted 'a room' instead of 'bedroom', like your solution was here - things would be different?It just feels like your logic is really flawed when faced with any scrutiny.
>>101843223wtf man I'm just trying to understand you
>>101843292Don't pretend to be me.>>101843223Remove "bed" from the negative of your prompt. You should get a bedroom, right?
>>101843292>Hey I'd like a room without a bed in it>How about a sitting room?>Yeah, that one's good... but... there's other rooms that doesn't have bed is it?>It is true, but fuck you I'm only gonna suggest a sitting room, fuck the other choices, I don't like creativity>Ok honey we're living
>>101840811Hello, I'm also retarded. I am downloading "flux1-dev-bnb-nf4.safetensors". I assume this is the correct file. Where do I put it?
>>101843310Sorry
I love how screencaps really look like random screencaps from some old movie or show
>>101843360in Stable Diffusion folder
>>101843403Thanks!
>>101843360me after dling comfy and the full model, realizing its not comfy, trying fp8 on flux and then finally getting the nf4. at least it works
>>101843481Same. I ended up deleting ComfyUI because it was so uncomfy.
>>101843310How about this?
>>101843524Yes, that works. I concede.
>>101843585https://www.youtube.com/watch?v=oXJvOJNUlCs
>>101843631
>>101843524can you please share your workflow
>>101843751Sure thing: https://files.catbox.moe/yjd2qq.png
>>101843825She's like the pink power ranger of her generation.
Dead chat?
How do you do speech bubbles and titles?
>>101844275ask for it
>>101844407in flux
> The VRAM requirement dropped from 24 to 12 GB relatively quickly and now, with bitsandbytes support and NF4, we are even looking at 8GB and possibly 6GB with a 3.5 to 4x inference speed boost.my 6GB is gonna gen so fucking hard. you just wait nerds.
>>101844822yeah it's now 8gb for the nf4 flux, but the fp16 t5xxl is still there, and that mf is 11gb big