As someone who constantly gave you shit for not having chapter markers, again thank you for adding them and staying consistent. It makes watching these significantly better.
Now humanity expect you to do the same for making videos about an open source Offline AI software that can generate both videos, image and audio including speech and singing at the same level of those AI website, and does not have any limitation of what we prompt we use and will generate what we write.
If this is truly v0.1, as in it's nowhere near done cooking, and it's already competitive with SD3, then I'm excited for how good Auraflow 1.0 is gonna be
I hope that this might become the new centralized main open source image ai. Maybe Pony and the civitAi community itself will focus on it so we have a new kong after the downfall of stability ai. Might take some time but if we guys want this to develop even better we have to talk about it and push it fruther to the Ai community so poeple dont keep sticking to the old outdated sd models
SD 1.5 is the only diffusion model that runs on my GTX 1060M with 6GB of vram. I'm glad that open source is pushing the boundaries of what's possible, but I'm not using any of these larger models until I upgrade my laptop, and I'm not doing that until my current one bites the dust.
Tried it out. Complex prompt with multiple characters and each with different clothing and color and holding different stuff is working. Normally on SDXL it mixes everything up. SD3 can seperate everything pretty well. So cool to see that Aura flow can seperate it as well !
No lie. Only reason SD1.5 is still the most used is because of the development in NSFW models. Even for non NSFW work, most of the work was done by people contributing to NSFW.
There are some things worth mentioning Aura flow is on v0.1. Can't even be compared to SD3. But I guess that the most important thing is: Open weights (sd3) vs open source (auraflow). Openweights disclose just that: the model, the weights. With auraflow you have EVERYTHING: from weights to the scripts used for training. that's a HUGE difference and will count a LOT for fine tunning by the community!
About their text prompts - I realised that when I add - with the exact letters - instead of just writing some text, in hyphens, Midjourney is capable of adding correctly written text into its generated image.
But Midjourney feels too Midjourneyesque after watching it. You can feel it is AI made most of the time because the shot is too beautiful cinematic détailled…
Hopefully the community updates their tools to work with Aura Flow soon enough. All this time with Stable Diffusion has shown that the best Image Gen Models is ultimately next useless without the tools to properly make use of them. Still I can't wait to see what Fine Tunes come out once this model is efficient enough for Medium to Lower power GPUs.
I was reading somewhere that the quality of captioning during training was the distinguishing factor in making the closed source models better. I think Pony's success has been due to better captioning, and it also seems that Aura Flow's efforts are paying off in the same way. I think SD3 was crippled by the fact that they tried to overly censor the base model because once people have the model there's no way to enforce NSFW checking after the fact. Dalle3 and MJ can have a less restrained base model because they ultimately can just not let the image go out if it's too much for their risk appetite.
Dall-E was free to choose any style it wanted because you didn't specify a style. So that result was just as accurate to the prompt as all the other models' results.
A fairer comparison of the tools would be to perform iterations on the results, say 5 or so, which is what it usually takes for me. It would take into account abilities like MJ to "paint" away glitches, enhance detail, etc. I'm more interested in what tool can produce the ultimate result than what's best on first shot.
Great video. Just wish you used a much more impresive variant of Stable Diffusion to really compare Aura to the other best open source option. Also, when it comes to the prompt, maybe we shouldn't look negatively upon a model using a cartoony style for example if we wanted it to be a realistic style but did not prompt that.
Even though I've been writing fiction with Microsoft Word or on paper since I was a little kid, I want a typewriter so badly. It's the middle ground between those two and something about it just feels authentic. I can type rather than write, but the writing is physically in front of me rather than on a screen.
That's probably the issue. At the beginning of the video you said open source had to rewrite its code "entirely", so of course we will have issues for a few months, this about the progress in 1 year from nightmare fueled spaghetti being eaten to what we have today on most paid platforms. It's near uncanny valley at this point. Making the program rewrite and retrain on everything will take just as long. Maybe a bit faster, but it still needs to be used and trained for month at least before it can come close to what it was before. The paid platforms never paused, never reset, and never had to retrain, so the gap between the two will continue to grow.
In the tests why did they all go for the Chrysler building style skyscraper? Is the Chrysler building really that well photographed that its poisons the outputs, or 'barges in' when skyscraper is mentioned.
I just saw some threads of people saying aura flow used ideogram images as dataset Edit: and I'm fine with that, if they are going to make their model open source, but the issue comes when they are using ai images to train an ai, if the ideogram images they used on the dataset have any type of errors aura flow will learn those errors as if they were correct
Tried and it can't do one thing better than the known image generators. There are more '2023 image generators' like Auraflow but I personally don't find that interesting. Thanks for the vid anyway.
Nice to see a new competitor. But the examples don't impress me,.. to be honest: They look all kind of boring and basic. Looking forward to the next iteration.
@MattVidPro You would not be wrong if you did "WebSim Wednesdays", where you make some lemon themed websites, or any websites, for 10+ minutes (or, more). Wednesdays would certainly prove to be your most viewed days, I'm sure - that is, until the Ai space gives us the next-new tool.
you can use it with comfy. at civitai you find the model and openart.ai/workflows/congdc/auraflow/gMFTu1GTMRPJf6AwW37V there is a workflow also at civitai, but dont know because the workflow says "training". both and probably more /different ones are easy to find by searching :) its not a new king, at least not yet ;) maybe it will be good, but for now its also slower than sdxl (around half the speed or less) and "only" a base model.
Using Ai on servers (non locally) is no more free and open than FB is. You're training and developing the software for them by using it and they are surely using every image you generate to train future models. If I were making anything original or creative I would only use a local generator.
I tested like 40-50 prompts with Auraflow and it was ALL terrible no matter which settings I used along with no difference between the location I used it from your links. So far even SDXL models give me better results. Auraflow needs way more fine tuning but I do see the potential since it’s open source. But even DALL-E was giving me better outputs for single subject basic Becky prompts too. Ideogram still reigns supreme for me as far as coherency and consistency vs the other closed models. I hope open source surpasses all these eventually.
Good to know I'm not crazy. Didn't test it as extensively, was just trying to get a simple logo for my website. It completely failed at getting the website name down, went to Ideogram and it was getting it it perfectly half the time.
Is it uncensored? no bad hand, bad face, bad body, bad fingers again? normal bone system supported(unlike openpose using a missing backbone skeleton) transparency background (PNG) supported? Constantly object generation?
for the text in midjourney you should have your text in double quotes not single quotes, hence why midjourney failed the test. not fair to midjourney imo, also i wish leonardo was included.
Well, your comparison with competitor should have compared with Stable Diffusion 1.5 and 2 or XL. SD is still the top open source one by far to compare with, just not SD 3.
20:23 using different image generators to achieve different styles is like buying a new phone every time the battery dies because you don't know they can be recharged. use... the prompt... to control... the style. Use the names of artists, or photographers, or publications, or countries, or architects, or camera lenses, or time periods, anything. Don't roll the dice on the style and mark down a model for how the dice landed, or even for how they usually land. tendencies are not capabilities! They are different! You would not believe the things I've seen SD 1.5 do with style prompting that it never even hints at knowing about with generic prompts. Who cares what style a model _usually_ chooses, if one model obeys your every request for a niche style, and another model acts like it can't hear you? I just want to see which of these models are willfully disobedient. Like Dall-E when you ask it to generate a vanilla minecraft screenshot and it applies shaders and sparkles and fog no matter what.
Doesn't seem to be good at rendering people (yet) - weird postures & wrong biology; hopefully they can improve those, so that this can beat SDXL for local generation.
Meh, not really impressed, especially with its size. I can get better results with custom SD 1.5 models and a good workflow on ComfyUI. Also, Aura Flow is huge, so most people will not be able to use it on their machine which is the best thing with open source models. Let's not forget, the future of image and video generation is with multimodal models that also have the intelligence of LLMs. Just the single modality of image generation will not be able to compete once GPT-5 and Claude 4 are released some time within the next 5-7 months. I just hope that Meta will release Llama 4 as an open source multimodal model, and that its smaller version will still be around 8-10b parameters together with image capabilities at DALL-E 3's level but with better understanding.
Nah bro we're not done with images gen, we really need it to be listining to prompts better and consistency , untrained brain/eye like yours wouldn't understand
Eh... it's alright. Could be useful for creating some scenes and then inpainting them as needed, but custom SDXL training and supplemental fine tuning with LoRas and PTI's is still the way to go if you want a model to do PRECISELY what you want it to do. Also, I can't get behind or support civitai. They allow FAR too much degeneracy on their platform and the fact that nobody even talks about it is WILD to me. I didn't even know wtf loliporn was until I started using that site. I'm just like what the FUCK am I looking at?! For real, they need to get their fucking shit together.
As someone who constantly gave you shit for not having chapter markers, again thank you for adding them and staying consistent. It makes watching these significantly better.
Commendable.
how?
Always has to be one person who is so shitty that makes the rest of us look better, thank you
Viewers ask and I do my best to accommodate. Been using AI to generate them from the transcript works decently well!
Now humanity expect you to do the same for making videos about an open source Offline AI software that can generate both videos, image and audio including speech and singing at the same level of those AI website, and does not have any limitation of what we prompt we use and will generate what we write.
If this is truly v0.1, as in it's nowhere near done cooking, and it's already competitive with SD3, then I'm excited for how good Auraflow 1.0 is gonna be
I hope that this might become the new centralized main open source image ai. Maybe Pony and the civitAi community itself will focus on it so we have a new kong after the downfall of stability ai. Might take some time but if we guys want this to develop even better we have to talk about it and push it fruther to the Ai community so poeple dont keep sticking to the old outdated sd models
We need a new Kong? What's wrong with King, Donkey, Diddy, Lank, Trixie, Cranky, and the others? 😂😂😂
SD 1.5 is the only diffusion model that runs on my GTX 1060M with 6GB of vram. I'm glad that open source is pushing the boundaries of what's possible, but I'm not using any of these larger models until I upgrade my laptop, and I'm not doing that until my current one bites the dust.
Tried it out. Complex prompt with multiple characters and each with different clothing and color and holding different stuff is working. Normally on SDXL it mixes everything up. SD3 can seperate everything pretty well. So cool to see that Aura flow can seperate it as well !
Is it process prompts as long as Dall-e-3?
Stable Diffusion 3 never fails to make me laugh. Whenever I'm having a bad day, all I have to do is generate a couple of SD3 images. 🤣
Just make sure it's NSFW friendly because that draws in the most innovation and support TBH.
No lie. Only reason SD1.5 is still the most used is because of the development in NSFW models. Even for non NSFW work, most of the work was done by people contributing to NSFW.
The internet is for porn
Very underrated channel. Thank you for the consistent updates, matt.
thank you!
266K subs. "Underrated"
Underrated?
@@StefanReich Yes, underrated.
@@universe6735Yes, underrated.
There are some things worth mentioning
Aura flow is on v0.1. Can't even be compared to SD3. But I guess that the most important thing is: Open weights (sd3) vs open source (auraflow). Openweights disclose just that: the model, the weights. With auraflow you have EVERYTHING: from weights to the scripts used for training. that's a HUGE difference and will count a LOT for fine tunning by the community!
Imagine the makers of auraflow knew he would test the lemon on the beach photo. So they finetuned it
If I was making one that's def what I would do lol
Based on the results I'm getting with AuraFlow it wouldn't surprise me.
Honestly the biggest + for me so far i got LOrAs i made for SD1.5 working with it out of the gate.
I stil use Stable diffusion XL and I think it is really great
Midjourney, their images are always the most beautiful, despite it's shortcomings.
I agree.
Lora will fix that. also paid lol.
About their text prompts - I realised that when I add - with the exact letters - instead of just writing some text, in hyphens, Midjourney is capable of adding correctly written text into its generated image.
But Midjourney feels too Midjourneyesque after watching it. You can feel it is AI made most of the time because the shot is too beautiful cinematic détailled…
The more competition, the better.
Hopefully the community updates their tools to work with Aura Flow soon enough.
All this time with Stable Diffusion has shown that the best Image Gen Models is ultimately next useless without the tools to properly make use of them.
Still I can't wait to see what Fine Tunes come out once this model is efficient enough for Medium to Lower power GPUs.
I was reading somewhere that the quality of captioning during training was the distinguishing factor in making the closed source models better. I think Pony's success has been due to better captioning, and it also seems that Aura Flow's efforts are paying off in the same way. I think SD3 was crippled by the fact that they tried to overly censor the base model because once people have the model there's no way to enforce NSFW checking after the fact. Dalle3 and MJ can have a less restrained base model because they ultimately can just not let the image go out if it's too much for their risk appetite.
If only we had SDXL comparison too to the mix though.
honestly stable cascade is better than all 3, doesnt matter unfortunately because base models suck anyways and no one is training it
I'm sure it can be much better as it's 6.8B model. Aura flow community have a lot of work to do to create good v.1.0 model.
Dall-E was free to choose any style it wanted because you didn't specify a style. So that result was just as accurate to the prompt as all the other models' results.
A fairer comparison of the tools would be to perform iterations on the results, say 5 or so, which is what it usually takes for me. It would take into account abilities like MJ to "paint" away glitches, enhance detail, etc. I'm more interested in what tool can produce the ultimate result than what's best on first shot.
Hello ! nice tests !! But, please : Draw us a summary table of your tests ! 😉
Thanks for the idea!
Aura you kidding?! Incredible!
LMAO
Nice comparison... Midjourney raw might give better generations. With raw off, it sometimes gets too creative.
Great video. Just wish you used a much more impresive variant of Stable Diffusion to really compare Aura to the other best open source option.
Also, when it comes to the prompt, maybe we shouldn't look negatively upon a model using a cartoony style for example if we wanted it to be a realistic style but did not prompt that.
Downloaded the 16GB model and it works in Auto1111 but the text is garbled. Any idea why this might be?
tested, like the others, crazy stuff coming out
Insane timing. I literally downloaded Fooocus last name..first time using stable diffusion and love it.
Fooocus gave me a lot of issues to run. Did you have any issues with getting it to run?
Last name?
@@grim789 No none at all. It was just clicking on the run file and it opened the website. No issues so far.
@@hipjoeroflmto4764 guessing he meant 'last night'.
11:30 It looks like one of those mobile game ads XD
5:15 Ideogram is my favorite also, especially with Text.
Even though I've been writing fiction with Microsoft Word or on paper since I was a little kid, I want a typewriter so badly. It's the middle ground between those two and something about it just feels authentic. I can type rather than write, but the writing is physically in front of me rather than on a screen.
imagine a gpt built into a typewritter and the type writter can auto type i bet it would seem like a ghost in the room lol
Nice to have more competition, especially open source, but so far Ideogram is far better for my prompts.
That's probably the issue. At the beginning of the video you said open source had to rewrite its code "entirely", so of course we will have issues for a few months, this about the progress in 1 year from nightmare fueled spaghetti being eaten to what we have today on most paid platforms. It's near uncanny valley at this point. Making the program rewrite and retrain on everything will take just as long. Maybe a bit faster, but it still needs to be used and trained for month at least before it can come close to what it was before. The paid platforms never paused, never reset, and never had to retrain, so the gap between the two will continue to grow.
Please don't ever stop making videos. lol Awesome content!
Someone just leaked Ideogram AI's source xd
Nice
In the tests why did they all go for the Chrysler building style skyscraper? Is the Chrysler building really that well photographed that its poisons the outputs, or 'barges in' when skyscraper is mentioned.
Thanks Matt. How can we install it locally?
Looking into this :)
Maybe pinokio has a local version.
Please cover Flux. Flux 1 DEV is the SD3 we deserved, and by elo also PASSES ultra.
This channel is truly pro AI.
I just saw some threads of people saying aura flow used ideogram images as dataset
Edit: and I'm fine with that, if they are going to make their model open source, but the issue comes when they are using ai images to train an ai, if the ideogram images they used on the dataset have any type of errors aura flow will learn those errors as if they were correct
Tried and it can't do one thing better than the known image generators. There are more '2023 image generators' like Auraflow but I personally don't find that interesting. Thanks for the vid anyway.
AuraFlow is trained on AI generated images and it shows, its output is way too artificial (or simply cartoony).
Matt!! Your finally back!!! Its been days!!! Its me kaylen
ummm so i tested it and my first generation was a person siting on a chair and they had no face and there feet were missing
My first six attempts - and each as bad as SD 2.0. Try any scene with "Joe Biden" and see.
For the typewriter prompt, Stable Diffusion and Ideogram were the only two that were smart enough to use actual typewriter font on the paper.
Actually facts
There's not many SD3 models on Civitai because it's banned
Any chance of you making a quick video on how to set Aura Flow up?
For sure going to look into it
@@MattVidPro Awesome, thank you!
cool, i didnt knew the backstory
Kind of wish you compared it to SDXL instead of SD3 since SDXL is generally considered to be the better model even though SD3 is newer.
Nice to see a new competitor. But the examples don't impress me,.. to be honest: They look all kind of boring and basic. Looking forward to the next iteration.
I hope to see this model on Poe
Tensor Art has quite a few SD3 models.
I'm not sure sd3 fine tunes can fix the mess, it's broken not in the way sd 1.5 was (just not great outputs)
@MattVidPro You would not be wrong if you did "WebSim Wednesdays", where you make some lemon themed websites, or any websites, for 10+ minutes (or, more).
Wednesdays would certainly prove to be your most viewed days, I'm sure - that is, until the Ai space gives us the next-new tool.
the demented happy look in the thumbnail is nothing like the the demented face while doing the video
The overall image impression of MJ is so much better though.
Kyutai on the other hand has still not released anything, even though they promised it.
Does/will it be usable locally, with ComfyUI or Automatic1111?
you can use it with comfy. at civitai you find the model and openart.ai/workflows/congdc/auraflow/gMFTu1GTMRPJf6AwW37V there is a workflow also at civitai, but dont know because the workflow says "training". both and probably more /different ones are easy to find by searching :)
its not a new king, at least not yet ;) maybe it will be good, but for now its also slower than sdxl (around half the speed or less) and "only" a base model.
sd3 has changed its rules to the better and also will release a updated version soon
Stable Diffusion: -1000 Aura / Aura Flow: +10000 Aura
does anyone know, how much Vram does this take?
How do you install it locally?
This is the right question. SD's ability to run locally is what truly sets it apart from the rest.
Using Ai on servers (non locally) is no more free and open than FB is. You're training and developing the software for them by using it and they are surely using every image you generate to train future models. If I were making anything original or creative I would only use a local generator.
I tested like 40-50 prompts with Auraflow and it was ALL terrible no matter which settings I used along with no difference between the location I used it from your links. So far even SDXL models give me better results. Auraflow needs way more fine tuning but I do see the potential since it’s open source. But even DALL-E was giving me better outputs for single subject basic Becky prompts too. Ideogram still reigns supreme for me as far as coherency and consistency vs the other closed models. I hope open source surpasses all these eventually.
Good to know I'm not crazy. Didn't test it as extensively, was just trying to get a simple logo for my website. It completely failed at getting the website name down, went to Ideogram and it was getting it it perfectly half the time.
Is it uncensored?
no bad hand, bad face, bad body, bad fingers again?
normal bone system supported(unlike openpose using a missing backbone skeleton)
transparency background (PNG) supported?
Constantly object generation?
Their website is full of bugs and the Aura AI is not that good even compared to SD3 in art comparison.
How can we run this locally?
8:48 ohh that Coca Cola logo there shouts for an law suit 😬
8:33 SD 3 seems terrible. I dont think ehstever you are running this in is doing a clip skip. Clip skip 2 is nearly required for all stsble diffusion.
Great vid, Matt. Thanks! But did you forget how to pronounce “aura”? 😂
I've grown more forgiving of various pronunciation over time... But checking YouGlish.... No one pronounces it the way Matt does here.
Matt is very young, he can't speak Latin.
hi matt! remember me from the livestream?
yes sir!
Wow! Thank you, Matt!!!
Nice 👍
When the world needed it the most, it showed up, unlike that bald kid with the tattoos.
What licence is it under?
Apache 2.0
@@MattVidPro So same as all the others. Would be interesting to have a true FOSS model under the GPL licence.
@@YeeLeeHawwhat is the problem with Apache licence?
Remember that there is no such thing as a piece of software that can "learn". It is still just ones and zeros and electrons moving in different ways.
All these models compared to DALL-E 3 and Midjourney are like babies fighting mike tyson.
ideogram fan boy. :)
Yessir i def am i use it the most
How big is this model in terms of parameters?
I believe around 7b params, maybe a bit more don’t quote me!!
@@MattVidPro "I believe around 7b params, maybe a bit more don’t quote me!!"
-MattVidProAI @@BackTiVi
I've heard it from Matt it's around 7b params
Whats needed for Local? Not interested in web usage.
*Hello Mat AI Need To Clarify Your Real Right ✅️ 😂*
for the text in midjourney you should have your text in double quotes not single quotes, hence why midjourney failed the test. not fair to midjourney imo, also i wish leonardo was included.
❤
Would like to see image to image
Well, your comparison with competitor should have compared with Stable Diffusion 1.5 and 2 or XL. SD is still the top open source one by far to compare with, just not SD 3.
20:23 using different image generators to achieve different styles is like buying a new phone every time the battery dies because you don't know they can be recharged.
use... the prompt... to control... the style.
Use the names of artists, or photographers, or publications, or countries, or architects, or camera lenses, or time periods, anything. Don't roll the dice on the style and mark down a model for how the dice landed, or even for how they usually land. tendencies are not capabilities! They are different! You would not believe the things I've seen SD 1.5 do with style prompting that it never even hints at knowing about with generic prompts.
Who cares what style a model _usually_ chooses, if one model obeys your every request for a niche style, and another model acts like it can't hear you? I just want to see which of these models are willfully disobedient. Like Dall-E when you ask it to generate a vanilla minecraft screenshot and it applies shaders and sparkles and fog no matter what.
what if we want to run it locally? (on my pc)?
If that's not the case then it can't be truly open source as they claim.
Doesn't seem to be good at rendering people (yet) - weird postures & wrong biology; hopefully they can improve those, so that this can beat SDXL for local generation.
Exactly!
I think Midjourney did pretty good overall. I smell bias.
Art is in the eye of the beholder. You guys are free to make your own thoughts after all
Will it become uncensored?
someone could make it uncensored because it is open source
+1000000 aura
Meh, not really impressed, especially with its size. I can get better results with custom SD 1.5 models and a good workflow on ComfyUI.
Also, Aura Flow is huge, so most people will not be able to use it on their machine which is the best thing with open source models.
Let's not forget, the future of image and video generation is with multimodal models that also have the intelligence of LLMs.
Just the single modality of image generation will not be able to compete once GPT-5 and Claude 4 are released some time within the next 5-7 months.
I just hope that Meta will release Llama 4 as an open source multimodal model, and that its smaller version will still be around 8-10b parameters together with image capabilities at DALL-E 3's level but with better understanding.
I don't really see the point of the title ... (looking at the content)
bro flux1 just out
Half the screen is black... Instead of not reading the prompt and criteria, use the dead space on the screen.
Yea, but image generation is so last year. Now it's all about videos.
Nah bro we're not done with images gen, we really need it to be listining to prompts better and consistency , untrained brain/eye like yours wouldn't understand
Eh... it's alright. Could be useful for creating some scenes and then inpainting them as needed, but custom SDXL training and supplemental fine tuning with LoRas and PTI's is still the way to go if you want a model to do PRECISELY what you want it to do. Also, I can't get behind or support civitai. They allow FAR too much degeneracy on their platform and the fact that nobody even talks about it is WILD to me. I didn't even know wtf loliporn was until I started using that site. I'm just like what the FUCK am I looking at?! For real, they need to get their fucking shit together.
I just unsubscribed because of this bullshit.
Same lol
Is ideogram your sponsor or something?
No
Ideogram is hands down the best.
I tried it and it thinks "young man" or "young woman" means child, so basically this AI is a grandpa.
I must be doing something wrong with Aura because the results are awful lol
The old king is dead, long live the king