AI Video Generation - General Discussion, Tips, Tricks, Frustrations and Showcases

[Casshern2]

The Technology forum can go for long stretches without much
activity at all. I tend to think that won't change any time soon,
especially with this thread.
- Casshern2​

It wasn't long after I started contributing to the Photorealistic AI Generated Images thread when I realized it really didn't quite belong in the JAV Discussion threads. But, like others, I kept posting along with subsequent members. As I started playing with AI video generation even while I'm in no way that great with image generation yet, I started posting examples. But I think I'm doing the right thing by moving that here. At least it's not in the way anymore in JAV Discussions.

I only know what I know so far in my first steps into this, and my lack of a great top tier modern PC is absolutely holding me back from any further potential with video generation, but I'm hoping there are others like me who may not have a golden rig but still want to see what can be done with available resources. Nothing says only the cool kids can have cool things. I'm hoping like the Images thread others will eventually post replies with their AI videos and/or their experiences generating on the web via a site/service or on their local machines. This is for general discussion, sharing tips and tricks and even venting because things aren't working out. I'm also hoping others will showcase things they've produced. We might all be able to learn a little of this and that together and end up helping each other get started with this. It's been fun for me so far.

I, myself, will try not to get long-winded (which will be a chore) or computer sounding as to not scare off any who land here. Things might get technical but hopefully presented in a non-intimidating way. At least, that's the goal. But, by all means, if anyone is pretty proficient with this already, lay it on us.

It's been a long day, so I'll leave this here. Please don't take this as one of those threads where the OP poses a question and just waits for others to start responding. Just didn't realize how tired I am as I type this. We've all been there. Next post will be my basic setup and the experience of trying things out.

 

[Casshern2]

I'm afraid I can only speak for one technology. LTX-Video. And even then, the local install of it, not the LTX-Studio available on their site. Reason being...I can generate what I want uncensored. Which should be the goal of anyone on Akiba. But I'm hoping members have had success with other model types and will share their eperiences.

First my setup (don't laugh). The beauty of LTX is that it is the best (that I've come across) at being able to generate something coherent on consumer grade systems. I had found articles on Reddit of people successfully using LTX on 10xx and 20xx cards. Not as fast, but it could work when using smaller models.

PC: Dell XPS 8500
OS: Windows 10 Home Edition (I know...I know...)
RAM: 32GB
Page File: 78.6 GB
Graphics Card: MSI Ventus GeForce RTX 3060 12GB GDDR6 PCI Express 4.0 RTX 3060 Ventus 2X 12G OC
[I started with a EVGA GeForce GTX 1060 3GB]

I had to use an internal power adaptor for the 3060 to work.


About models, that's regarding the LTX-Video models. There are a number that have come out. I'm actually one generation behind but that's me. This is where it gets intimidating for some. Best thing to do is try what you'd like to use then work your way down to what you can comfortably use. Unfortunately, these models are pretty large, so the trial and error (I found) can be lengthy.

Here are some LTX models I've tried:

ltxv-13b-0.9.8-distilled-fp8.safetensors - 14.6 GB
ltxv-13b-0.9.7-distilled-fp8.safetensors - 14.6 GB (currently using)
ltx-video-2b-v0.9.5.safetensors - 5.90 GB
ltx-video-2b-v0.9.1.safetensors - 5.32 GB
ltx-video-2b-v0.9.safetensors - 8.72 GB
(there are quantized models that are smaller...but you'll have to Google that, I haven't tried)

The 13b stands for "13 billion paramaters" and of course 2b is the "2 billion paramanters" model. File size comparison says it all, one is trained a hell of a lot more than the other. I started with the 2b models when I still had the GTX 1060, but since I've been on the 13b. Really nothing else of interest to explain about 13b vs 2b, but Google it if you're curious.

For it all to come together, though, you will need something to control it. The go-to I found almost everyone using is the very visually intimidating ComfyUI. But, it's not that bad at all...untill you set your eyes on the Workflows people are creating. Here is glimpse of the one I've been having fun with.