The institutions are so deftly corrupt and incompetent they banned 10s videogen and bugfix AI.

I agree with you but here (at least for me) it's not about "which loss has it better" but it's an existence proof that "pure videogen pixels models got this too", directly disproving claims of the contrary. For this, size doesn't matter.

Is there a way to use Higgsfield MCP connector with other videogen models?

A crucial aspect is that the VAEs used to train the videogen models perform no-better than chance, but the diffusion models then figure it out themselves. This is in contrast with VideoMAE and V-JEPA, which learn some representations of physics through their encoders !

To be clear, this is not a V-JEPA or VideoMAE diss, just resurrecting the fact that "pure videogen" models may indeed learn an explicit model of the world/physics as a byproduct. Also cc @mapo1 we chatted about this and you also intuitively pushed back against such claim.

cant video edit in my region unfortunately. will do for simple video gen tho! few weeks ago it wasn't really better than seedance 2 in videogen

If videogen models implicitly learn object dynamics, could multi-agent interaction dynamics (coordination, communication, pursuit/evasion) also emerge without being explicitly modeled?

Because that's what you prompted the videogen model for???

Isn't that what he expected from VideoGen models? We know that ImageGen models learn, among other things, depth and segmentation. They have to have some kind of world representation.

@sama well look at that: You abort poor Sora, harr Codex pivot (fafo cannibals eh, @Disney) to *catch up* to @DarioAmodei (his IPO gonna tag yrs taintstain) now (unexpectedly) meine CEO just approved scaling up ai videogen (since 2023 here, scroll. back) which justifies. flops..

At the beginning of the year, there was another paper who did this check only for V-MAE and V-JEPA, and showed both have some understanding of physics. This new paper essentially extends the study to diffusion-based "pure videogen" models and shows they understand physics very well.

You may have recently heard claims that video generation models are "dumb" about physics, and only "world models" (V-JEPA, specifically) have a valid internal model of physics. This turns out to be false. In a recent paper, researchers show that a LINEAR probe of diffusion videogen models predict various "physics" very well, significantly better than V-JEPA or VideoMAE (and plain VAE just sucks). This is noteworthy, because a *linear* probe being this accurate shows that the model has a pretty explicit internal representation of the physics!

I'm not really being pedantic no, yes you do imply you think that it is good because of course you do, so do I, that's why I said it was a self interest argument. What was missing was an appeal to something else, which would make it not a self interest argument. Open source does not just mean a race to the bottom on price, it means a race to the bottom on everything, like alignment or refusals. Do you really think the world is better off with open source image and videogen? easy access to revenge porn and misinformation generation? etc.? They do collaborate, on alignment especially, which is what matters most. Sharing capabilities would do nothing w.r.t. pro-socialness. They do these things because they also, rightly, believe a multipolar future would be dystopian.

Nice work! I'm just curious about what's the difference between sparse videogen team and quant video team🧐😋