A lot of folks in this thread have mentioned that the problem with the current generation of models is that only 1 in (?) prompts returns something useful. Isn't that exactly what a reward model is supposed to help improve? I'm not an ML person by any means so the entire concept of reward models feels like creating something from nothing, so very curious to understand more.
Which AI's allow you to keep training the model? Most are pre-trained without you knowing how. If you use an open source LLM, you could probably do it, but then you already need to have a lot more understanding of it, and be more technical, and have proper hardware. Most AI's I have seen and worked with don't have an option to keep training it. You just use the model as-is, possibly with a initial prompt to tell the AI in what kind of fashion it should respond.
I'm not so sure how much that is relevant to Meta Movie Gen. I've tried all the tools: Luma, Runway, Kling
Luma is by far the worst and relatively compared to Runway and Kling by far produces the worst quality and unstable video. Runway has that distinctive "photo in the foreground with animated background" signature that turns many off.
Kling and Runway share that same "picture stability" issue that is rampant requiring several prompts before getting something usable (note I don't even include Luma because its output just isn't competitive imho).
This Meta Movie Gen seems to make heavy usage of SAM2 model which gets me super excited as I've always thought that would bring about that spatiotemporal golden chalice we always wanted, evident by the prompt based editing and tracking of objects in the scene (incredible achievement btw).
Until I have the tool ready to try I will withhold any prejudgements but from my own personal experiences with generative video, this Meta movie gen is quite possibly SOTA.
I simply have not seen this level of stability and confidence in output. Resolutional quality aside (which already Kling and Runway are at top of the game), the sheer amount of training data that Meta must have at disposal must be far more than what Kling (scrapes almost the entirety of Western content, copyrights be damned) and Runway can ever hope to acquire, plus the top notch talented researchers and deep learning experts they house and feed, makes me very optimistic that Meta and/or Google will achieve SOTA across the board.
Microsoft on the other hand has been puttering along by going all in on OpenAI (above, below and beside) which has been largely disappointing in terms of deliverability and performance and trying to stifle competition and protect its feeble economic moat via the recently failed regulatory capture attempt.
TLDR: this is quite possibly SOTA and Meta/Google have far more training data then anybody in the existing space. Luma is trash.
