Just a few months ago, realtime LTX-2.3 for Flipbook was barely doable, and only with a massive hit to quality. I have no affiliation but I am blown away by what @fal just did with H3 Max. An uncompromising OOM shift of the frontier bigger than anything else I’ve ever seen. They took the best open-weights video model and somehow made it both better than every other model, closed and open, and 15x faster than the next best thing. faster than realtime!! Biggest update for me is that it breaks this tight tradeoff we've come to expect between quality, speed, and cost. Apparently you CAN improve all 3 at once, and by a lot. A similar thing just happened in LLM land with GLM-5.3, but I realize now it's actually possible someone could come along soon and go further? take a model like that, but actually make it even better than Fable AND 15x cheaper AND faster? What happens then? Also, in general for media gen, a threshold has been crossed here. In theory we could replace all content/software/video games, etc. with on demand video generation now. Some richer inputs/conditioning would be nice, but we now have photorealistic audio/video with text rendering, etc. generating faster than realtime for any subject you can imagine. It just costs $140/hr (for now)