Training Wan2.1 LoRAs

[hila-chefer]

Hi :)

We are trying to train our own Wan2.1 LoRAs that are compatible with the code on diffusers, but we couldn't find a training script in the code.
Can anyone please direct us at the training script used for the LoRAs on diffusers?

Also, does anyone know the compute requirements for Wan2.1 14B vs 1.3B and how long the training take for each case?

Thanks in advance!

[a-r-r-o-w]

Hi Hila 👋

We're maintaining training scripts at https://github.com/a-r-r-o-w/finetrainers for Wan. It's not as hackable as the single file scripts but hopefully something to start with.

For 1.3b with dp_degree 8, each step takes 1-10s depending on resolution/number of frames.

For 14b with dp_degree 2 and dp_shards 4, it takes anywhere from 5-100s depending on resolution/frames.

We're working on context parallel + FA2/FA3 to speedup training but it's a WIP with other priorities.

(excited for whatever your team is going to create haha!)

[hila-chefer]

Thank you so much for your quick response and for your tremendous contribution to the open source community @a-r-r-o-w ❤️🙏

[ZhouQianang]

@a-r-r-o-w Wan-14b? 5–100 seconds? How can I do that? I really need it.