Because coders only use the latest technologies available, without any prior optimization. For example, there's an optimized version of Stable Diffusion that can use older video cards, not just the latest ones. And the result is the same as with newer video cards. Of course, it is faster with newer video cards, but the result remains the same.
I have followed the instructions and started a test, but had to interrupt. The time this takes to produce video is so much, even on top-end hardware, that it is not practical for production. This should be mentioned on the video. Also, renting a computer on the cloud, even at 8 cents per hour or whatever would amount to quite a sum for even a modest production, and would take too long. This is just not practical, except for producing a video for the views and getting revenue from the ads. Will not suscribe. I have done some calculations to show how slow and/or inefficient this process is, even on a 4090. Look at the stats on his video at 30:30: almost 4 secs per iteration, and the process uses 12x40 iterations per process... that's 1920 secs or 32 minutes for a 7 sec video (sample audio 1). Imagine you want a presenter for a standard 15 minute video. It would take 68 HOURS to generate: almost 3 days. On the cloud he suggests, that's 54USD. You can see that this can mount up pretty fast, and that's without even counting if you have to re-render due to mistakes or whatever. So, this is the problem with channels that want to hype and bring enthusiasm to some technology: they are too biased. 30:30 for the part where he minimally mentions this core problem. This is NOT personal computing: it belongs to huge datacenters consuming hundreds of thousands of watts where developers can manipulate the parameters and see results in real time, the way Microsoft demoed his technology. Imagine: to make it real time, you would need the equivalent computing capacity of 275 of those RTX 4090, plus the infrastructure to keep them running. That's like 137,500 Watts per hour, or a 90 Horse Power, 1.8 cc SOHC engine running at full capacity and burning some 9 gal of gasoline per hour, just so you han have your little window with a face moving at the dance of your mouse and keyboard.
Thank you for being honest about the ones that look weird. For some reason, a lot of RU-vidrs are applauding these crappy looking pictures and videos.??? I like the honesty
If only he had been honest about how slow and/or inefficient this process is, even on a 4090. Look at the stats on his video at 30:30: almost 4 secs per iteration, and the process uses 12x40 iterations per process... that's 1920 secs or 32 minutes for a 7 sec video (sample audio 1). Imagine you want a presenter for a standard 15 minute video. It would take 68 HOURS to generate: almost 3 days. On the cloud he suggests, that's 54USD. You can see that this can mount up pretty fast, and that's without even counting if you have to re-render due to mistakes or whatever. So, this is the problem with channels that want to hype and bring enthusiasm to some technology: they are too biased. 30:30 for the part where he minimally mentions this core problem. This is NOT personal computing: it belongs to huge datacenters consuming hundreds of thousands of watts where developers can manipulate the parameters and see results in real time, the way Microsoft demoed his technology. Imagine: to make it real time, you would need the equivalent computing capacity of 275 of those RTX 4090, plus the infrastructure to keep them running. That's like 137,500 Watts per hour, or a 90 Horse Power, 1.8 cc SOHC engine running at full capacity and burning some 9 gal of gasoline per hour, just so you han have your little window with a face moving at the dance of your mouse and keyboard.
@@I.Am.Nobody yeah, for sure as there is a lot of vested interest in saving costs. I personally have a friend who worked for a couple of years in optimizing AI tensors to the current nVidia architecture, using specific numeric types that could best be processed by these new units... but it IS going to take a lot of time, and we should be real about that.
Sounds amazing. This combined with some nice embeddings could make a great conventional bot. Over time you could generate more answers to similar questions and make the conversation sound a lot more natural. I am currently building a coding curse and having an AI Tutor to ask questions that actually looks like a human would be great.
Nice!! Agents will have a face generated image but this Avatar will be animated Not by simple user prompts but in Real Time with Zero Latency by an LLM behind this Avatar and those same *Mimic Outputs Signals* can be used to feed the *Inputs PWM Signals* of electric motors to animate a Fully Autonomous Robotic Physical World Version, with these Physical Avataras performing as the Puppet of these Artificial Neural Networks. That's The *Last Stage* of this tech during the *AGI Incarnation World Event* of 2028... And So The Ray Kurzweil Singularity begins.
Looks like an interesting alternative to Hedra. I like that it is more configurable. Future models should maybe have an option to describe the character and take this into account when doing the animation. For example age, or if it's an anime or some kind of doll.
Im not really sure if i should be happy or terrified that this is closed source. Those who have full access can do insanely good deepfakes. Propaganda will reach a whole new level.
I went in this panic in February 2023. I realized the internet is on a risk to die. I archived lots of movies, documentaries, lectures, went on eBay and ordered encyclopedias for all fields, just in case. Helped me be more relaxed working with AI and accepting the journey
@@theAIsearch I think in court it will probably be not too hard to figure out if it's AI generated, by proffesionals who can examine it (but who knows, maybe people will find a way to make it hard to prove, we'll have to wait and see) But it will have huge consequenses on social media in general Most people already don't read beyond the headlines. So even if it will be debunked as AI generated, the damage will already be done So yeah. The potential of this tech is terrifying
1. The tech is in early stages and developing very quickly. Eventually it will not be obvious at all 2. Many are already fooled by obvious false social media posts.
I've been using Vidnoz to do deep fakes of streamers and send the video to them in discord. Never fails to get a laugh! You can make them say what ever you want? Ever wanted to be a VIP or a mod in chat? Now you can make a fake video of the streamer promoting you! KEKW
The only drawback to this technology concerns the eyes and head movements. When you're talking to someone or looking at a video camera, cell phone, etc., your eyes aren't staring blearily at the lens all the time, or even at a person, all the time due to our process of thoughts. Even our face moves when we talk. Only robots, or AI, do not do that. But coders are not psychologists anyway.
Man I installed everything and followed every instruction, but upon trying to make a video, it takes me a very long time to render. I use an RTX 3080 8GB GDDR6 VRAM. What Can I do to speed up the rendering. Now it takes like 7 hours to make a video, help me please.
How do you see the open source alternatives comparing to the copyrighted ones? Are they getting close? I think that with open source, we could not only deal better with AI ethics ("Skynet", "big Brother", etc.), but we could get better results, since AI needs lot of data and open source helps with accessibility = it's free. I really appreciate your work btw :D
open source is as good or only a bit behind. eg. Hallo is as good as Hedra, which is the best closed model you can use right now. i think Meta is also releasing an open source competitor to GPT-4o which will be natively multimodal. expect open source to only be a few months behind
@@theAIsearch@johnangel5480 He's probably using a conda environment and that's a big issue for those who don't want to install python system-wide but rather in the conda sandbox. Search for "How can I activate a Conda environment from PowerShell?" in Stack Overflow.
Am I taking crazy pills or what this thing ran for like two and a half hours and then bottomed out and said it wants to restart now if I just paid for 2 hours of servers?
Absolutely crazy and beautiful. Imagine a utopia where there are no degenerates hungry for power or malicious intentions. Humanity, being kind, would achieve greatness in such a short time.
@@Derpyzilla894 It's not about 'image' generation, you need to see further. There are many powerful models out there, probably some really close to AGI, maybe from Anthropic or OpenAI or some national research private labs. The point is, if we were in a utopia, those models would already be out there. But you can't let an AGI roam free with so many mentally unstable people, asking how to create explosives, making deepfakes of people in NSFW contexts; the list goes on. That's why many AIs are so restricted.
@@Derpyzilla894 It's not about 'image' generation. There are many powerful models out there, probably some really close to AGI, maybe from Anthropic or OpenAI or some national research private labs. The point is, if we were in a utopia, those models would already be out there. But you can't let an AGI roam free with so many mentally unstable people, asking how to create explosives, making deepfakes of people in NSFW contexts; the list goes on. That's why many AIs are so restricted.
While the installation, a red error message appears says: ERROR: Package 'hallo' requires a different Python: 3.9.5 not in '=3.10' Although I checked the Python version, and it is 3.10.11 .. Why that?
Hm..... More and more I look at these new models I feel like someone is definitely gonna slap a model into bot and let em learn continuous and wallah we made the first ever humanoid.
Well I tried this with an actual gradio interface that someone was kind enough to send me the link and for right now it's just too slow and when I say slow I mean really slow, for a 2 minute video it was going on 5 hours when I finally shut it down. I hope at some point they optimize it because it looks interesting
when installing install ps1 I get errors it seems like things arent installing right but it says its done at the bottom in pwrshell and there is no torch even when using venv
