We had a recent update to DynamiCrafter ComfyUI wrapper, this time resolving the Conditional Image Leakage (CIL) issue. The authors for CIL are still working on the non-watermarked model and various resolutions. So I had spent some time trying out the watermarked model for 576x1024 resolution.

I used 945M, disabled analytic_init, 26 steps, cfg 7.5.
Generation speed is slower (obviously due to additional computations).

Video (1.5s):
https://imgur.com/a/o3zdDuD

I managed to get the character to turn her head to face the viewer.
Now I know this is not Kling, Luma, Gen3 standard, but at least this is what our open source research community has managed to progress.

For now, I am going to check if they release the non-watermarked model everyday as I am done waiting for 14 hours for Luma to process my generation request.