[Question] Clarification on 'sparse_ratio' and 'kv_ratio': Impact on Quality vs. Speed

[raymond-etet]

Hi there,

I am currently using FlashVSR for video super-resolution and tuning the inference configuration. I observed a trade-off between speed and parameter settings, and I would like to confirm the best practices for achieving the highest possible quality.

My Observation:
When I increased the sparse_ratio from 4.5 to 10.0:

• The inference speed increased from 1.03 it/s to 1.15 it/s (on a 60s video).
• The speed improved, but I am concerned about the potential loss in reconstruction quality.

My Questions:
Since my priority is Quality > Speed, could you please clarify the following:

1. sparse_ratio: Does a higher value mean higher sparsity (less computation) and therefore lower quality? Should I decrease this value (e.g., to < 4.5) to get better details?
2. kv_ratio: Similarly, for kv_ratio (currently testing 3.0), does a smaller value result in better quality?
3. Recommendation: If I want the best possible visual result and do not care about inference time, what range of values would you recommend for these two parameters?

Thank you for your hard work on this project!