feat: add community works LeMiCa

[joelulu]

Thanks for the great work！

LeMiCa - LeMiCa: A training-free caching acceleration framework proposed by the Data Science & Artificial Intelligence Research Institute of China Unicom, featuring DAG-based globally optimal modeling. LeMiCa offers adjustable speedup options for HunyuanVideo-1.5, delivering up to 2.9× (T2V) and 3.9× (I2V) faster inference. See LeMiCa4HunyuanVideo1.5 for details.

🧩 Text-to-Video (T2V) H100x4

Model	HunyuanVideo-1.5 (latency min)	LeMiCa (B=25)	LeMiCa (B=20)	LeMiCa (B=15)
T2V 720p	8.98	4.84 (1.85x)	4.03 (2.23x)	3.14 (2.86x)

HunyuanVideo1.5_T2V_720P.mp4

---

🧩 Image-to-Video (I2V) H100x4

Model	HunyuanVideo-1.5 (latency min)	LeMiCa (B=25)	LeMiCa (B=20)	LeMiCa (B=15)	LeMiCa (B=10)
I2V 720p	9.10	4.92 (1.85x)	4.04 (2.25x)	3.17 (2.87x)	2.35 (3.88x)

HunyuanVideo1.5_I2V_720P.mp4

Official Support

Moreover, the acceleration paths provided by LeMiCa can be quickly adapted and seamlessly integrated into the HunyuanVideo-1.5 framework (e.g., see infer_state.py#L48). We warmly welcome official support and would be excited to see this contribution merged upstream.

[joelulu]

I2V looks much faster than T2V mainly because B=10 causes larger content shifts in T2V, so we recommend B=15. With B=15 (as in the table), T2V and I2V have nearly the same runtime and speedup.