Best Tools and extensions for STABLE DIFFUSION AI art

[@guilhermegamer]

FIRST! LoL. I am From Brazil . I tought you would do no more videos. Gonna watch it now. Already liked it.

[@Not4Talent_AI]

hahhahah thank you so much!! I'll continue to do videos, it's just that I have very little time to do so and I take too long xD

[@Not4Talent_AI]

List of extensions and tools YOU suggested: Thanks to everyone that collaborates! -Snixs: Infinite image browsing it allows you to see all your generated images in a gallery inside SD and you can see the prompts and settings you used for each image and of course send it to img2img etc. you can also search by a certain keyword or model you used and more. github.com/zanllp/sd-webui-infinite-image-browsing An app from Civitai called SD Prompt Reader which also allows you yo read prompts and settings of a certain image without opening SD. github.com/receyuki/stable-diffusion-prompt-reader -Rasukix: civitai helper extension, auto downloads the image overlay so you don't have to read each lora/checkpoint looking for the right one github.com/butaixianran/Stable-Diffusion-Webui-Civitai-Helper -It's Eloii 🐌: For my landscape I use Gaea from Quadspinner to create realistic terrain quadspinner.com/ (A tutorial I found kzbin.info/www/bejne/nIOlgo2deMtgmbs&ab_channel=theAstropath ) -Leonard Petit-Breuilh: For posing I use EasyPose on iOS/Android/Steam (Has some cool variations, maybe someone needs to pose on their phone. Here it is) -István Nagy: The Umprompted extension has something similar to Inpaint Anything, where you can define a mask using a prompt (txt2mask), then use it for a second img2img pass. github.com/ThereforeGames/unprompted (This has so much other stuff it could go with it's own tutorial. Here is the manual) github.com/ThereforeGames/unprompted/blob/main/docs/MANUAL.md -Karearea: Also just wanted to say, when it comes to photobashing, you can use generated images too! If you’re worried about using another artist’s work, you can generate your backgrounds, characters, etc and stitch them all together for a base image -DarkStoorM: 1- Tag Autocomplete extension This is a really helpful tool for autocompleting tags in your prompt. This extension was mostly made for Booru tags, but you can disable those and use regular autocomplete, since booru tags don't really work on regular models (mostly). The best thing about this is, that you can autocomplete embeddings, loras and hypernets (requires checking out the extension settings!!!). To disable the default booru tag autocompletion, you can change the "tag filename" from default and use your own csv, for example if you only want Enhancers or different Quality tags only as danbooru csv is basically just for nsfw. It's super useful for Embeddings/Loras, because you can just type left angle bracket and a part of the name, like

[@losing_interest_in_everything]

Additionnal ressources : kzbin.info/www/bejne/jpe0dGSFrdmoeqs kzbin.info/www/bejne/aJbFg4pvbpmSrLM

[@XxFennasxX]

I really am enjoying regional prompter right now. It can set prompts to apply to specific subsections within an image.

[@Not4Talent_AI]

It's great tbh, very useful. I usually use latent couple (pretty much the same thing). Very nice if you dont want to go inpainting hahaha

[@jeremycointin1996]

Good video. Love the fast ( no fat ) pace. I use a lot of these for the art on my channel. By the way ... did you see Tafi? THAT is gonna be the future for shizzle!!!

[@Not4Talent_AI]

thanks!! I saw, even though it is, (I read from the community of Olivio Sarikas), a feature for DAZ 3D. It doesn't create models, it just plays with the preexisting settings on there. Or something like that xD The closest thing to real text to 3d models is LatentLabs, and it isnt good enough for much yet

Very cool selection, thanks! That Inpaint Anything seems interesting, I'll give it a go. The Umprompted extension has something similar, where you can define a mask using a prompt (txt2mask), then use it for a second img2img pass.

[@Not4Talent_AI]

Oh, really? I'll check it out in a moment and pin it then, thanks!

[@Not4Talent_AI]

do you know if this is the one? github.com/ThereforeGames/unprompted I cant find the mask option you said. Maybe it is another one or they don't talk about it here

@@Not4Talent_AI Yes it is. The txt2mask in the "Stable Diffusion Shortcodes" section at the bottom, you have to expand it.

[@Not4Talent_AI]

oh, my bad I'm stupid XD Thank you! I'll pin it then :3 It looks like a really complex extension. Ty!

[@losing_interest_in_everything]

For my landscape I use Gaea from Quadspinner to create realistic terrain

[@Not4Talent_AI]

uuuuuu it looks nice! It is for download but looks helpful, I'll pin it. Thanks!

[@jamesmiller9924]

This is excellent, even if it moves a bit quick. I have a question, though, and my question is about photobashing. How do you prompt for that? So you've got two characters in your "paladins fighting" image, and they're basically the same in substance (knight swinging a sword or whatever). But what if you've got two characters doing different things? One person holding out a hand in a "stop there" pose, and the other with their hands at their sides looking startled? Once you have your photobashed reference image in Img2img, how do you prompt in a way that SD can make variations without totally messing it up? Just keep the denoising low enough?

[@Not4Talent_AI]

tyty!!! for the first one, you could either inpaint or do something like what I do in this video: kzbin.info/www/bejne/oZfScmOujNtgZrs And for the second one, you could either go with low denoise, or IP 2 p + CN tile. Even just CN tile regularly with some denoise (depends on your image)

[@sickvr7680]

Que buenos vídeos que haces pibe, te felicito, suscripto!

[@Not4Talent_AI]

muchas gracias!

[@morizanova]

Super and awesome Resources for harnessing AI Image generator in God Mode Power Level ! Soemthing which Lex Luthor always wanted from Superman . Sharing the knowledges 😅

[@Not4Talent_AI]

hahhahahhaha, glad it helped!! tyty

[@USBEN.]

Very well done video. Keep it up.

[@Not4Talent_AI]

thank you!!

[@zhoudraconis7526]

What interface is this that you are using?

[@Not4Talent_AI]

in this video I use a lot of stuff. but if you are referring to the Interface I use to generate images, its Stable Diffusion automatic 1111 local install. 100% free, uncensored and customizable to an extent. You just need a decent gpu

[@nikgrid]

Excellent really informative video! thanks.

[@Not4Talent_AI]

Thank you! Happy that it helped!

[@pastuh]

daz3d is nice, but obvious company will bancrupt if they not introduce AI tools inside app...

[@Not4Talent_AI]

Who knows XD I think they are adding them though. Not sure if it is inside the app or not. but basically their tool is a text to 3D that uses their assets to generate the input

[@faultline3936]

I wonder if they'll make it as a premium feature, since most of the good sht is always locked behind a paywall lol Edit: Was talking about PoseMyArt

[@Not4Talent_AI]

Not sure about what part specifically you are talking about, but lets hope not xD Even if they do, other websites will appear that make it free. probably xd

[@faultline3936]

@@Not4Talent_AI Completely forgot about that lmao I was referring to PoseMyArt

[@Not4Talent_AI]

@@faultline3936 Oh ahhahahahha Figured it would be, but couldn't assume xD I don't think they will. but maybe, now that they seem aware that SD users like the website, they'll create some payed options regarding that.

[@joeskis]

so basically "AI" still requires you to do a ton of work to get what you want.

[@Not4Talent_AI]

Yup😂😂😂😂

[@BZAKether]

Excellent video, as always. I am redirecting anyone who asks me about SD to your channel.

[@Not4Talent_AI]

thank you very much!!!!

[@Karearearea]

Also just wanted to say, when it comes to photobashing, you can use generated images too! If you’re worried about using another artist’s work, you can generate your backgrounds, characters, etc and stitch them all together for a base image

[@Not4Talent_AI]

actually good point, always forget to mention that. Amma pin this if you dont mind. Thanks!

[@Karearearea]

@@Not4Talent_AI Yeah, no worries! Also thanks for all your work including links and timestamps and everything, it’s an easily overlooked thing but it makes following along SO much easier

[@Not4Talent_AI]

@@Karearearea thank you for watching and the feedback+tips!!

[@jibcot8541]

The Ultimate SD Upscale script is a must have for high res images, especially for people without much Vram.

[@DJVARAO]

The best content for SD is coming from professional artists. Thank you!

[@Not4Talent_AI]

Thank you so much for the kind comment! Hope it helped :D

[@ninjuhdelic]

holy shit!?!?!?! working swords??????

[@Not4Talent_AI]

😂😂

[@eddiemauro.design]

Sos lo maximo. Tremendo duro del explanation 😎

[@Not4Talent_AI]

hahahahahha gracias!!

[@ArielTavori]

Mixamo makes me sad. I'm glad they haven't taken it down (yet), but it feels like abandonware. I don't think the features or assets have been updated in like ~5 years or so..? Bummer. It's still the best tool of its kind, but such a waste of potential...

[@Not4Talent_AI]

Yeah... it has been like that for so long. Don't really know why. They probably didn't find a way to monetize the site, or it just wasn't worth it. even though I don't think it would be that hard.

[@ARXeno_GT]

How can we get the pose to make the image

[@Not4Talent_AI]

Like, the programm to follow the pose or how to get the pose reference de program needs? For the first. Anywhere. Random images and then using a prepricessor. Posemyart. Blender. Op extensions... etc For the second, you need to install controlnet with the openpose model

[@Nik.leonard]

For posing I use EasyPose on iOS/Android/Steam but I guess PoseMyArt is better (?).

[@Not4Talent_AI]

Better depends on what one needs. This looks like it has some cool variations and stuff. Aside from being from mobile, that can be helpful to someone, I guess XD. Pinned!

[@baihu5914]

Can I ask if there a way to see other artists, prompt they using in their art ? Some of them are so good ,

[@Not4Talent_AI]

There is, using "png info" if you have the metadata of the image. There are times where people have inpainted or done other changes to the image so it might not be correct. In civit AI people post their prompts too

[@Rasukix]

civitai helper extension, auto downloads the image overlay so you don't have to read each lora/checkpoint looking for the right one haha

[@Not4Talent_AI]

cool! ty!

[@lefourbe5596]

oh ? nice one yep.

[@FrancoisDuplan]

Hey! Totally digging your tutorials - they're a blast and super easy to get. Found your channel on Stable Diffusion just yesterday, and you're already my top pick. Any chance you're thinking about a step-by-step on making our own LoRAs? With your ace way of explaining stuff, it would be really nice. Especially as there aren't really any comprehensive, easy-to-follow tutorials on the subject. Keep up the good work !

[@Not4Talent_AI]

Thank you so much! Super happy to hear that! Im planning on making one eventually, but I don't know that much of it. Not enough to really make a guide that teches stuff people dont already teach. Seeing how many of you are interested, I should probably move it up my priority list. Next video is a pretty hard one to make. So maybe in the meantime I can think and read about LORA training a bit more. (Wont be able to expermient much cuz it takes a ton of time tho😅) It will come, Ill push it sooner than I intended seeing the demand. Thank you very much!

[@FrancoisDuplan]

@@Not4Talent_AI That's awesome to hear! :) I understand that you want to make sure you have enough knowledge on the subject before making a tutorial. I'll be eagerly awaiting that LoRA tutorial whenever you're ready to make it. Good luck with your next video !

[@Not4Talent_AI]

@@FrancoisDuplan thank you so much!! Helps a lot🧡

[@anitaart6501]

来自北京。感谢你的分享，受益匪浅，特别感谢。加油

[@Not4Talent_AI]

Ty for watching !

[@spencereaston8292]

When an image renders then disappears, like what you saw in the inpaint anything, take a look at the terminal. It will say that it can't write to some folder. For me it was /tmp/gradio. I manually created it and that solved that issue.

[@Not4Talent_AI]

oh, I'll check that, thank you so much!!

[@cesarsantos854]

Now I am expecting text to pose tools.

[@Not4Talent_AI]

that would be awesome ngl. And then to animation hahahha

[@Mente_Fugaz]

I wonder what will be humans making after that lol... Not being able to do nothing by ouselves is boring tbh

[@Not4Talent_AI]

@@Mente_Fugaz hahhahha yeah, that is a pretty big concern with AI. mainly with stuff like chatGPT. Amma forget how to count in 3 years

[@Mente_Fugaz]

@@Not4Talent_AI yeah.. in my opinion the mere fact that is not valuable or worth to learn to draw by yourself anymore is a defeat by itself... because we loose authenticity and admiration of the talent Hopefully this tools will keep developing like tools for people to use instead of just replacements of what your tutorials shows. The fun thing with art is to make it by ouserlves after all hehe

[@Not4Talent_AI]

@@Mente_Fugaz I pretty much agree. I do think that these are just tools aimed to help. That's why I focus mainly on pure control over a result. Because that means YOU where the one who thought what they wanted to represent, and brought it to life using these tools. I don't think it will replace anything, but it will 100% increase the supply without increasing the demand. Learning to draw is still really fun though, but now you don't need to know a lot of the tedious stuff. (even though, if you know it you will be able to stand out more. but that is just for now, we don't know where AI will be in 3 years). I personally, as a 3D artist that was learning how to draw when AI popped out, think that AI is just a tool that will allow for people to create what they wanted to create but couldn't. Which I ultimately see as a good thing. But with the expense of changing the whole ART ecosystem with that. (damn this was long)

[@abstrack97]

Your content is lowkey underrated. Keep the extensions showcase coming!

[@Not4Talent_AI]

Thank you so much!! Will do ones this ones get obsolete I guess hahhaha

[@DarkStoorM_]

Speaking of faster, there is something, that really needs to be pointed out apart from extensions (I will mention an extension at the end of this text wall :D), which is more of a Quality of Life improvement. A couple things, actually: (warning, I don't know how will it behave on other GPUs/systems, so it might not work as expected on all PCs or slower hardware, I only tested it on rtx4090, as I don't have my old PC anymore) *Speeding up the generation time/iterations* 1) (important) If you are on Torch 2.0 with latest CUDNN (you can download separate binaries from Nvidia Developer website, but I can't compare them to the built-in cudnn) - and with AUTOMATIC1111 1.3.~, there is a settings option to change the Cross Attention Optimization. Enable your --xformer flag as usual in your .bat file, but in the settings, change your Cross Attention Optimization to "sdp - scaled dot product". This has literally no impact on the resulting image and it speeds up the generation time by almost 80% (for me at least). People say it affects prompt editing, but there is literally no difference in complex prompts. 2) (not really necessary) If you really like very frequent Live Preview updates on bigger amount of steps, *do not use Full quality previews* if you have something like 500ms update every 0/1 steps period, even with beefy GPUs. It will delay your SD every update, because it will send a full quality image to your UI, which will make you wait for additional step every time the live preview updates. For example, on 0step updates and 500ms, your SD will in fact be quick, but this will slow you down every 4 progress updates or every 500ms window - it depends on your hardware. A solution to this is switching to *TAESD* from Full, which is a lower quality preview, but it does not block your progress, still allowing to see _somewhat_ good preview on full speed. *This setting is good if you only need to see the approximate result composition.* If you don't really care, something from 2000-3000ms is completely fine on full quality. 3) *If you are only generating 512x512 images when looking for bases to upscale or inpainting* and plan on generating a bigger batch, *never exceed 75 tokens* on either positive or negative prompt. This will cut your iterations down by almost a half. This size is kind of a middle ground for generation times if your prompts fit into one block (75 tokens). This will have lesser effect on bigger images and no impact on batched generations. This Kind of matters if you mix negative embeddings, like NG_DeepNegative, which already is 70 tokens long. NOTE: This does not affect DDIM/PLMS samplers For example, from my tests: - < 75 tokens on 256x256: 42~ it/s - > 75 tokens: 22.08 it/s - < 75 tokens on 512x512: 32~ it/s - > 75 tokens: 20.08~ it/s - < 75 tokens on 768x768: 15.5 it/s - > 75 tokens: 13.40 it/s - < 75 tokens on 1024x1024: 7.70 it/s - > 75 tokens: 6.90 it/s On 512x768 size it's 23.5 it/s vs 14.5 it/s, so it still matters 4) *Shared GPU usage* - this one is kind of obvious, if you run any software, that actively uses your GPU, it will also slow down the generation. I like having Twitch on my other monitor, but just noticed that's +10% GPU usage for me with actual -10% SD performance. Fun fact: when you watch youtube and have the video controls active, like when you hover over the player, the GPU usage will be way higher, impacting your performance :) --------------- *Tag Autocomplete extension* I don't know if you already covered this extension in the past, but this is a really helpful tool for autocompleting tags in your prompt. This extension was mostly made for Booru tags, *but* you can disable those and use regular autocomplete, since booru tags don't really work on regular models (mostly). The best thing about this is, that you can autocomplete embeddings, loras and hypernets (requires checking out the extension settings!!!). To disable the default booru tag autocompletion, you can change the "tag filename" from default and use your own csv, for example if you only want Enhancers or different Quality tags only as danbooru csv is basically just for nsfw. It's super useful for Embeddings/Loras, because you can just type left angle bracket and a part of the name, like

[@Not4Talent_AI]

Holly s..., you thank you so much for your time typing this!! It is really useful information I really had not heard (read) anywhere before. If this video dind't have a foced pinned comment, this would be the one. Hope people give like to this bc it deserves to be seen. (another option is commenting it as a response on the main pinned comment). But still, thank you very much for sharing this info. I don't really know how to check if I'm running on Torch 2.0 with the latest cuDNN, but that is a me problem cuz I'm dumb af xD I'll pin the extension part if you don't mind! I hadn't covered it before. Knew about it but never really tried it. I just now, thanks to you, learned that it can autocomplete loras and embeddings, which is pretty nice ngl.

[@DarkStoorM_]

@@quercus3290 There is no VRAM difference. Currently, I'm inpainting on 1280x1280 pieces with two CNet units cached, 9.7GB VRAM usage on idle, 13.3GB when generating. It stays the same for both SDP and xformers. xformers takes 40 seconds while SDP takes 13 seconds. This is even more huge when inpainting

[@DarkStoorM_]

@@Not4Talent_AI You can scroll at the very bottom of the webui, the info is in the footer. I'm currently on "torch: 2.0.1+cu118 "

[@Not4Talent_AI]

@@DarkStoorM_ oh sht, thanks!!! I'll pin this too if it is that easy to see, thank you!

[@DarkStoorM_]

@@quercus3290 Oops, youtube removed my comment because of the link - but TLDR: I'm on rtx4090 and with your sample I got 24s vs 8s generation time

[@Karearearea]

These are such amazing guides, thank you

[@Not4Talent_AI]

Thank you!! Happy to hear they help :D

[@itsalwaysme123]

Super killer video as always. Surprised you haven't blown up yet, but I have no doubts you'll get there!

[@Not4Talent_AI]

Thank you!! I would consider having 2k subs in such a short time span "blowing up" hahahha The support has been mind-blowing. ty all so much!