Can LightX2V be combined with FastWan for near-realtime performance on H100?

[tonyabracadabra]

Hi team,

I’m trying to maximize performance for WAN video generation on an NVIDIA H100, with the goal of getting as close to realtime inference as possible.

I’ve been exploring both:

• FastWan (sparse-distilled WAN checkpoints that natively run in 3–4 steps), and
• LightX2V (Lightning LoRAs + runtime optimizations like FlashAttention, quantization, self-forcing, etc.).

I’ve seen some community discussions where people experimented with applying LightX2V Lightning LoRAs on top of FastWan checkpoints, claiming they could drop to 2 steps. The results seem mixed, especially for motion quality.

A few questions:

1. Is it technically valid to combine FastWan with LightX2V Lightning for extra performance, or are the Lightning LoRAs strictly trained for base WAN2.1/2.2 models?
2. On an H100, do you expect such a combination could realistically achieve sub-second per frame performance (or near-realtime for 5s 480p/720p clips)?
3. If quality degradation is expected, are there recommended configs (LoRA rank/strength, sampler settings) that would minimize the trade-offs?

Lower quality is acceptable in my use case — I mainly care about pushing latency down as far as possible on H100. Just wondering if LightX2V + FastWan could be the right path, or if they’re intended to be parallel solutions used separately.

Thanks for your work on this project — really excited to see how far WAN can be pushed!