Questions about inference VRAM, sparse-attention/distilled weights, and lightweight variants

Hi, thanks for releasing this great work.

I have a few questions regarding inference efficiency and future release plans:

1. Could you please share the approximate peak GPU memory requirement for running AnyRecon inference with the released setting? For example, what is the expected peak VRAM for generating a 40-frame sequence at the resolution used in the paper?

2. In the paper, the efficient version combines 4-step diffusion distillation with sparse attention. May I ask whether there is an estimated timeline for releasing the 4-step distilled model with sparse attention?

3. Given the high cost of the 14B bidirectional backbone, I wonder whether you have considered lighter variants, such as smaller Wan backbones or streaming/autoregressive designs for long-trajectory reconstruction.

Any insights or suggestions would be greatly appreciated. Thank you again for your excellent work and for making the project available to the community.

Best regards

No file, test, or entry point is named. Start by reviewing the released inference setting and the paper's efficient-version description; the issue would be complete only when the requested VRAM estimate and plans for distilled, sparse-attention, or lightweight variants are documented.