Reduce the denoising steps.

[mingzhe11]

I'd like to ask, how can I reduce the number of steps? I see that the default denoising steps are 50.
It almost costs 10 minutes to transfer a video with 60 VRAM.

I'd like to ask if it's possible to accelerate LightX2V LoRa or anything else to reduce the runtime.
Look forward to your reply.
Thanks a lot.

[yxbian23]

I'd like to ask, how can I reduce the number of steps? I see that the default denoising steps are 50. It almost costs 10 minutes to transfer a video with 60 VRAM.

I'd like to ask if it's possible to accelerate LightX2V LoRa or anything else to reduce the runtime. Look forward to your reply. Thanks a lot.

Hi @mingzhe11,
Thanks for your question!
Since Video-As-Prompt is built on top of Wan2.1 and CogVideoX's DiT architecture, acceleration techniques that work for those base models should generally be applicable to ours as well.

Although I haven't had a chance to thoroughly test this myself yet, you could try lowering num_inference_steps in the pipeline call — the base models typically work reasonably well with 30 steps.

Additionally, DiffSynth-Studio has integrated our model and includes various acceleration optimizations. You may find their implementation helpful as a reference for speeding up inference.
Hope this helps! Feel free to share your findings if you discover any effective acceleration methods.