Is any recipe to make camera more stabilized

[vird]

~30-50% of my video generations damaged by at least one big change of camera angle
(I see that there is something with camera control in roadmap)
Is there some low effort recipe to fix that?

I've already tried inject in prompt Stabilized camera. Slow cinematic camera movement. linear camera motion. and negative prompt augmented with Camera shaking. Camera makes quick moves. Camera changes angle rapidly. That makes at least not 50-90% of video damaged as default. But it's still too bad to wait 40 minutes and then got not too good result

[ymcki]

I too get this problem for video2world diffusion but it seems like text2world diffusion doesn't have this problem.

For video2world diffusion, I find that if I try prompts from different vision model, e.g. llama-3.2-vision, sometimes the camera become more stable.

Of course, this is not an ideal solution. Let's see if others have better ideas.

[jonstephens85]

I also notice this on video2world. After about 1 second, the camera wants to look around.

[jpenningCA]

If you're using text2world, you could post-train using a dataset of stable camera videos? I haven't tested this myself, but may be worth a test? https://github.com/NVIDIA/Cosmos/blob/main/cosmos1/models/POST_TRAINING.md