Check out my courses AI Art courses. Midjourney Mastery www.udemy.com/course/midjourney-mastery/?referralCode=92BFBB305B81A1C7D1A0 Create and Sell AI Art www.udemy.com/course/make-and-sell-ai-art/?referralCode=C67A8D247B2578D4E762
But SD 1.5 has negative prompts already. Used them on Lexica, Hugging Face, and Replicate. While Stability's own Dream Studio still doesn't have them even in the 2.0 model.
From experience, you have to give SD (no matter the version) very long and detailed prompt to get close to Midjourney. Midjourney runs on a cloud, so the prompts can be short but still output great images because it already preselects popular styles and flavors. I still prefer SD, as you have so much more creative control over your images.
Agree. SD requires some ideas & artistry whereas MJ seems like a plug-&-play for people that don't have the ideas or commitment to create their own style.
Thanks for revisiting SD 2 - it got a lot of negative press when it first launched, I think partly because they pushed it out with great fanfare but little guiidance on how to best use it. But actually, it's OK. And still open, so we may see some interesting community enhancements in near future.
you missing the real point, SD2 needs specific models more then negative prompts, but trust me if i say that with a specific well done model and e solid negative prompt, the results are waaay better then MJ4 , especially cause of freedom you got on SD2, just my 2 cents.
Midjourney look is too defined and limited, is very hard to go out of it. MD Is a more driven for people without ideas and original aesthetic style, SD is more allaround creative tool.
Awesome! I have trying to get coloring pages via SD but they come out pretty disfigured and poor quality. Could you please help me to get decent coloring book-style images in SD? Thanks again
I love both SD and MJ. I typically use both of them. MJ for the original creation, SD for in-painting fixes and maybe in the future I'll also use SD more for up-scaling too, I've not really explored that enough yet, I've been using ciaiNNer for my up-scaling needs. (plus also some Photoshop work at the end to clean things up)
Hands will be perfected at some point in the future if they keep on developing this tech, it mainly is a question of when rather than if or so I believe.
It's not just about negative prompts more than you generally seem to have to be very specific and make sure certain keywords are in the standard prompt too. As an example, I trained a model on a flat art style via hypernetwork, yet whenever I just used the tag, it would still render stuff in photorealism, even with "photography" in the negative prompt. I wasn't until I specified it was a flat art illustration that it actually came through.
It says in the documentation for SD 2.0 that every generated image has a watermark embedded into it. I am suspicious as to wether thiis its whats causing the wierd distortion in images without neg prompts. In so many of my generated images im noticing the same exact wavy/ripple pattern all over generated pictures.
Yeah I can't go back to MidJourney after having all the options and controls in Stable Diffusion. The end result may not be as magical. But that is what Photoshop is for to finalize the image anyway.
There is actually a model that was trained for stable diffusion from midjourney images. Its called openjourney. It can create midjourney like images when you use that model
better as in closer to human art when drawing humans or characters? then neither is good. zoom in on the generated images. both AIs often draw deformed hands/fingers or generate weird artifacts at random places. I can name at least one thing that makes me suspect that the image is AI-generated for each image shown in this video.
We need more involved guidance concerning negative prompts. Which ones are useless, how many can you add before the whole thing collapses in on itself, which ones are too powerful. Been trying to get SD to do werewolves but I can't get it to maky any that are anywhere close to what MJ can make, but MJs censorship is royally killing me. Feels like I'm stuck using the worse of both worlds.
There isn't sufficient(or consistent) training data on "werewolf", so it just thinks it's "human making a weird face". Your best bet is to make an embed for that token.
I was wondering has anyone tried prompts that described scenes from wars or the people suffering from the ravages of war? I wonder what image the algorithm will come up with if we asked "5 year old Turkish boy lying face down restless on a sandy white beach washing up on to the shores after a shipwreck with no known next of kin"
So the VAE is the part which takes your prompt and converts it into an embedding to create the parameters for the model. So... could SD2.0 be improved with just a better VAE?
Using both, mostly SD for Inpainting, but still :p Although, you're talking about n****y in SD2.0, didn't they basically remove it from their "main" model ? Or are you talking about other models?
Yeah, they also added a digital watermark to every rendered image too and is what I suspect is the reason for the images being so bad without negative prompts.
Porn was removed from SD 2.0. Not sure about nudity and partial clothing but from my own results it looks like naked legs and waists were removed too. They should go all the way down to the conservative rabbit hole and exclude anything not _halal_ in Islam. Such as long hair and ankles.
SD 2.0 big nope. I'm not talking about absence of porn and 'greg rutkowski style'. The new model is an idiot about whole body anatomy (dressed and partially dressed). Remarkably worse than 1.5. Oddly enough, 2.0 spits out distorted nudity where 1.5 produces partially dressed images. Negative prompts are brilliant but we've already had them in SD 1.5 (try them on Lexica, Hugging Face, and Replicate), while they are still missing in Stability's own Dream Studio even in the 2.0 model.
P.S. Stability AI shouldn't stop at excluding naked legs and waists. They should go all the way down to the conservative rabbit hole and exclude anything not _halal_ in Islam. Such as hair and ankles. Because, why not? Is Western morality superior to Eastern?
You are talking about SD in general, not about SD 2.0. Clarify next time, you telepath. Which is possibly wrong anyway, because you haven't tried fine-tuned flavours of SD. Have a taste of it free in Finetuned Diffusion at Hugging Face.
I want to experiment with 2.0 & see what I can manage with it, but I’m still blown away by the difference SD 1.5 made over 1.4. Especially if you update to GFPGAN 1.4 instead of the old 1.3 or 1.2 models for fixing faces (or start using CodeFormer instead or supplementally), & update your VAE file if it’s out of date in addition to the main .ckpt StableDiffusion model… Cumulatively, between those upgrades & just learning how to optimize my settings & workflow, the quality of art I was able to produce skyrocketed soon after. Portraits & landscapes especially look phenomenal. It’s cool if negative prompts can compensate for some of the big downsides 2.0 seems to have, but it still seems like in many respects a step backwards (just not universally, because the new CLIP encoder seems capable of interpreting more abstract or semantic concepts & more complex phraseology better than the old one)… but when it comes to just a simple portrait like some of the realistic ones you compared here, SD 1.5 is clearly capable of outperforming 2.0 without negative prompts… Which makes me wonder, if you used the same combo of positive & negative prompts in 1.5, how would it compare against 2.0 & Midjourney 4? Would the negative prompts take it to an even higher height above & beyond 2.0, or do the negative prompts actually have such a more beneficial effect on 2.0 thanks to the new CLIP model that they make the difference & place 2.0 ahead. I mean, just the ease of use of not having to go through all of that negative prompting as a necessity just to get tolerable quality is a big, big pro in 1.5’s favor. So if it’s additionally capable of being refined to even better results if you take the time to add negative ones in, then that would really seal it for me that there’s very little benefit to 2.0. But yeah, I’d just discourage people from making overall judgements of Midjourney vs StableDiffusion based on the somewhat unfair comparison of Midjourney 4 (a clear most recent step forward) against StableDiffusion 2.0 (in many respects a step backwards). Whenever they release a more refined new version that is an unambiguous improvement over 1.5, I suspect that will rival Midjourney 4- but even as a huge fan of SD & it’s free & open-source project model, I’m not so biased that I’m going to pretend Midjourney 4 isn’t obviously an amazing advancement for Midjourney! Especially with more stylized & conceptual artwork, involving symbols or associations that SD struggles with, it’s amazing what I’ve seen people produce with it. Btw if every dumb ass billionaire with a cult of personality from Elon Musk to Jeff Bezos is going to own a generative AI algorithm, from Dall-E to Imagen, I do not look forward to Kanye’s generative AI that exclusively produces anti-Semitic tropes. 😒 lol…
What?.. StableDiffusion is free & open-source dawg. That’s one of the things that it has over Midjourney or Dall-E, that makes it unique & great. The idea of holding it against a free, not-for-profit algorithm that they aren’t raking in money, as if that was a downside, is completely missing the point.
MJ still beats SD in neg prompted every picture you showed... They are using Discord feedback to improve the algorithm. And they are still using real Artists work, which should be illegal under the IP protection laws.
![[Full Episode] The Restaurant War Thailand ศึกพ่อค้าซ่าแม่ค้าแซ่บ Episode 2 | 29 ก.ย. 67](http://i.ytimg.com/vi/r3lfLNH6Ke4/mqdefault.jpg)
Check out my courses AI Art courses.
Midjourney Mastery
www.udemy.com/course/midjourney-mastery/?referralCode=92BFBB305B81A1C7D1A0
Create and Sell AI Art
www.udemy.com/course/make-and-sell-ai-art/?referralCode=C67A8D247B2578D4E762
"Stable Diffusion 2 needs negative prompts" is I think the most accurate statement I have heard all year
But SD 1.5 has negative prompts already. Used them on Lexica, Hugging Face, and Replicate. While Stability's own Dream Studio still doesn't have them even in the 2.0 model.
Emad speaking the truth for a change? Right
From experience, you have to give SD (no matter the version) very long and detailed prompt to get close to Midjourney.
Midjourney runs on a cloud, so the prompts can be short but still output great images because it already preselects popular styles and flavors. I still prefer SD, as you have so much more creative control over your images.
pretty annoying business, I say.
Agree. SD requires some ideas & artistry whereas MJ seems like a plug-&-play for people that don't have the ideas or commitment to create their own style.
Thanks for highlighting the importance of negative prompts.
You are welcome!
Thanks for revisiting SD 2 - it got a lot of negative press when it first launched, I think partly because they pushed it out with great fanfare but little guiidance on how to best use it.
But actually, it's OK. And still open, so we may see some interesting community enhancements in near future.
Totally agree
I think with embeddings sd is already better, and it’s so easy to use the embeddings… the only think I need is the mix mode from mid journey now
What do you wrote to get the embedding? I can't seem to figure it out.
you missing the real point, SD2 needs specific models more then negative prompts, but trust me if i say that with a specific well done model and e solid negative prompt, the results are waaay better then MJ4 , especially cause of freedom you got on SD2, just my 2 cents.
How do you use this models.i have embeddings but can't seem to figure how to use them.
Midjourney look is too defined and limited, is very hard to go out of it. MD Is a more driven for people without ideas and original aesthetic style, SD is more allaround creative tool.
Total. Game. Changer.
Thanks.
I was already going to put in the trash SD 2.0, but this changed everything, and it works.
Awesome! I have trying to get coloring pages via SD but they come out pretty disfigured and poor quality. Could you please help me to get decent coloring book-style images in SD? Thanks again
Let me check
Thank you for putting negative prompts in description
Midjourney looks much better. It's sad they haven't improved SD much and now there's a need for negative prompts too.
I love both SD and MJ. I typically use both of them. MJ for the original creation, SD for in-painting fixes and maybe in the future I'll also use SD more for up-scaling too, I've not really explored that enough yet, I've been using ciaiNNer for my up-scaling needs. (plus also some Photoshop work at the end to clean things up)
I love Michael Jackson too.
I agree, this makes a lot of sense. I’m using a similar work flow
Use both SD and MJ together and you get it all.... ...except good hands, we seemingly will never get good hands :) Great vid friend.
Thanks 👍
hands are hard for artists too, but I bet this will be a solved problem in like a year.
Hands will be perfected at some point in the future if they keep on developing this tech, it mainly is a question of when rather than if or so I believe.
Thank you for the video and thank you for both listening and responding to your audience. Keep up the good work! Are you using Automatic1111?
Yes I am!
He _is_ Automatic1111.
Generally I like the SD outputs more. They seem to be crispier without that dreamy MD look. Then again I like Ernie better than both
It's not just about negative prompts more than you generally seem to have to be very specific and make sure certain keywords are in the standard prompt too. As an example, I trained a model on a flat art style via hypernetwork, yet whenever I just used the tag, it would still render stuff in photorealism, even with "photography" in the negative prompt. I wasn't until I specified it was a flat art illustration that it actually came through.
Hey Samson 🙌 What a pleasant surprise to get your video as a recommendation
I'm glad to reconnect in the TH-cam realm!
Thanks for dropping by and hope to see you again
It says in the documentation for SD 2.0 that every generated image has a watermark embedded into it. I am suspicious as to wether thiis its whats causing the wierd distortion in images without neg prompts. In so many of my generated images im noticing the same exact wavy/ripple pattern all over generated pictures.
I have not seen that - could you link the docs?
Nope. And you can disable it
Maaaaaan, youre such a legend Sampson 👏
Thanks Koi Boi ☺️!
a portrait of a fairy a luminous dress, mouth closed, long hair, wind, sky, clouds, the moon, moonlight, stars, universe, fireflies, butterflies, lights, lens flares effects, swirly bokeh, brush effect, concept art, celestial, amazing, astonishing, wonderful, beautiful, highly detailed, centered, digital art
disfigured, kitsch, ugly, oversaturated, greain, low-res, Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, , mangled, old, surreal, big nose , low-quality
Poorly drawn hands. My generations hands are drawn with extra fingers... (facepalm)
Are you using Image Prompts in all these generations, or just text?
Some had image prompts most didn’t
Yeah I can't go back to MidJourney after having all the options and controls in Stable Diffusion. The end result may not be as magical. But that is what Photoshop is for to finalize the image anyway.
There is actually a model that was trained for stable diffusion from midjourney images. Its called openjourney. It can create midjourney like images when you use that model
better as in closer to human art when drawing humans or characters? then neither is good. zoom in on the generated images. both AIs often draw deformed hands/fingers or generate weird artifacts at random places. I can name at least one thing that makes me suspect that the image is AI-generated for each image shown in this video.
MJ is still better imo, it takes so much more work in SD. Not to mention their ethics are better at MJ.
from where stable diffusion 2.o but i use automatic 1111 with ckpt
We need more involved guidance concerning negative prompts.
Which ones are useless, how many can you add before the whole thing collapses in on itself, which ones are too powerful.
Been trying to get SD to do werewolves but I can't get it to maky any that are anywhere close to what MJ can make, but MJs censorship is royally killing me.
Feels like I'm stuck using the worse of both worlds.
There isn't sufficient(or consistent) training data on "werewolf", so it just thinks it's "human making a weird face". Your best bet is to make an embed for that token.
Great follow up video!
Thank you!
Subscribed.
This is how a comparison video should be done
Why is nobody have done a proper Hand model yet ???✋🖐👐
Can I get some prompts for those
got it! stable diffusion is better for deep faking
I was wondering has anyone tried prompts that described scenes from wars or the people suffering from the ravages of war? I wonder what image the algorithm will come up with if we asked "5 year old Turkish boy lying face down restless on a sandy white beach washing up on to the shores after a shipwreck with no known next of kin"
😅
what's stopping you??
So the VAE is the part which takes your prompt and converts it into an embedding to create the parameters for the model. So... could SD2.0 be improved with just a better VAE?
stable diffusion 2 outputs very weird looking shadows on images
ohhk i'd disagree with this, mj v4 blows stable diffusion out the water, with or without negative prompts, its not even close these days tbh
That's the case now...but SD might very well catch up with MJ and will get as good as MJ is in the future, anything is still open and possible.
Using both, mostly SD for Inpainting, but still :p
Although, you're talking about n****y in SD2.0, didn't they basically remove it from their "main" model ? Or are you talking about other models?
Yeah, they also added a digital watermark to every rendered image too and is what I suspect is the reason for the images being so bad without negative prompts.
@@pelvist SD has always had a digital watermark. It should be imperceptible.
Porn was removed from SD 2.0. Not sure about nudity and partial clothing but from my own results it looks like naked legs and waists were removed too. They should go all the way down to the conservative rabbit hole and exclude anything not _halal_ in Islam. Such as long hair and ankles.
SD 2.0 big nope. I'm not talking about absence of porn and 'greg rutkowski style'. The new model is an idiot about whole body anatomy (dressed and partially dressed). Remarkably worse than 1.5. Oddly enough, 2.0 spits out distorted nudity where 1.5 produces partially dressed images.
Negative prompts are brilliant but we've already had them in SD 1.5 (try them on Lexica, Hugging Face, and Replicate), while they are still missing in Stability's own Dream Studio even in the 2.0 model.
P.S. Stability AI shouldn't stop at excluding naked legs and waists. They should go all the way down to the conservative rabbit hole and exclude anything not _halal_ in Islam. Such as hair and ankles. Because, why not? Is Western morality superior to Eastern?
right but it's so bad when it comes to digital art
You are talking about SD in general, not about SD 2.0. Clarify next time, you telepath.
Which is possibly wrong anyway, because you haven't tried fine-tuned flavours of SD. Have a taste of it free in Finetuned Diffusion at Hugging Face.
@@brexitgreens I'm talking exclusively about SD 2.0, it doesn't have artists and prompting digital art yields bad results
@@brexitgreens I'll give a try to finetuned though
@@brexitgreens I tried it, it's cool but it lacks samplers
I'm truly surprised! Great video!
Glad you liked it!
Nice follow up to yesterday's video!
Hope you enjoyed it!
I want to experiment with 2.0 & see what I can manage with it, but I’m still blown away by the difference SD 1.5 made over 1.4. Especially if you update to GFPGAN 1.4 instead of the old 1.3 or 1.2 models for fixing faces (or start using CodeFormer instead or supplementally), & update your VAE file if it’s out of date in addition to the main .ckpt StableDiffusion model… Cumulatively, between those upgrades & just learning how to optimize my settings & workflow, the quality of art I was able to produce skyrocketed soon after. Portraits & landscapes especially look phenomenal. It’s cool if negative prompts can compensate for some of the big downsides 2.0 seems to have, but it still seems like in many respects a step backwards (just not universally, because the new CLIP encoder seems capable of interpreting more abstract or semantic concepts & more complex phraseology better than the old one)… but when it comes to just a simple portrait like some of the realistic ones you compared here, SD 1.5 is clearly capable of outperforming 2.0 without negative prompts… Which makes me wonder, if you used the same combo of positive & negative prompts in 1.5, how would it compare against 2.0 & Midjourney 4? Would the negative prompts take it to an even higher height above & beyond 2.0, or do the negative prompts actually have such a more beneficial effect on 2.0 thanks to the new CLIP model that they make the difference & place 2.0 ahead. I mean, just the ease of use of not having to go through all of that negative prompting as a necessity just to get tolerable quality is a big, big pro in 1.5’s favor. So if it’s additionally capable of being refined to even better results if you take the time to add negative ones in, then that would really seal it for me that there’s very little benefit to 2.0. But yeah, I’d just discourage people from making overall judgements of Midjourney vs StableDiffusion based on the somewhat unfair comparison of Midjourney 4 (a clear most recent step forward) against StableDiffusion 2.0 (in many respects a step backwards).
Whenever they release a more refined new version that is an unambiguous improvement over 1.5, I suspect that will rival Midjourney 4- but even as a huge fan of SD & it’s free & open-source project model, I’m not so biased that I’m going to pretend Midjourney 4 isn’t obviously an amazing advancement for Midjourney! Especially with more stylized & conceptual artwork, involving symbols or associations that SD struggles with, it’s amazing what I’ve seen people produce with it.
Btw if every dumb ass billionaire with a cult of personality from Elon Musk to Jeff Bezos is going to own a generative AI algorithm, from Dall-E to Imagen, I do not look forward to Kanye’s generative AI that exclusively produces anti-Semitic tropes. 😒 lol…
thank you!! :D
You're welcome!
Money talks. Midjourney is raking in the money. They can afford the greatest minds.
What?.. StableDiffusion is free & open-source dawg. That’s one of the things that it has over Midjourney or Dall-E, that makes it unique & great. The idea of holding it against a free, not-for-profit algorithm that they aren’t raking in money, as if that was a downside, is completely missing the point.
Thankyou!
You're welcome!
Creating a good positive and negative prompt is an art in on itself, quite difficult... so everytime i get a good result is by pure accident.
"almost"
...there is no variation on reality where midjourney isnt a cheap monetized copy
Can you run SD on Mac? I’ve only tried it with Google Colab but I heard you need the paid version now.
Yeah I have it running on my M1
This boils down to: can you have Nvidia in your Mac?
P.S. There are no free Macs. Even on Google Colab. So you might pay for it as well.
MJ still beats SD in neg prompted every picture you showed...
They are using Discord feedback to improve the algorithm. And they are still using real Artists work, which should be illegal under the IP protection laws.
No it shouldn't
Thanks for the video.
waiting for your reply on email.