Feature request: LoRA / adapter support for SANA-Streaming (v2v live inference)

First, congrats on this great model! I'm using SANA-Streaming through Reactor and on my A100, and it opens up a lot of scenarios that were impossible before!

I'd like to train small style/domain LoRAs (e.g. rain, overcast, night) and load them into the streaming v2v inference — ideally swappable per session. Three specifics:

1. Training — is LoRA (or SFT-LoRA) training supported for the v2v / streaming model, not just the image model? If so, which script/config, and what dataset format? (docs/sana_cosmos_rl.md mentions sana-video-sft-lora via cosmos-rl — does that target the streaming v2v net?)
2. Loading — the live inference does build_model(...) + load_state_dict(strict=True) on a single .pth, with no adapter path. How would I load a LoRA adapter into that, or is the intended route to merge the LoRA into the base checkpoint and load the merged .pth?
3. Roadmap — if it's not supported yet, is v2v LoRA planned?

Don't hesitate to come back to me if you have any questions.

Thanks in advance,

Pierre