Wan 2.2 Scene Change LoRA by MQ Lab

This LoRA is a game-changer in WAN video generation!

While its basic function is to force scene changes, it also has another special feature:

It enables the use of T2V LoRA in I2V workflows.

By using this LoRA and T2V LoRA together in an I2V workflow, you can recreate T2V LoRA scenes while preserving the characters in the input image. A standard T2V LoRA model is fine, but it needs to be well-trained. If you'd like to try it, please use the T2V LoRA model I've uploaded. It's designed to be used in conjunction with Scene Change LoRA.

Prompt ex

The screen goes dark and then lights up again to show the new scene.

{your prompt}

T2V LoRA training method

Tool : Ostris AI-Toolkit

Model : Wan2.2 14B(T2V)

Resolution : 512 x 512

Rank : 32

Steps : around 2000

Number of videos in the dataset : around 20

*The number needed to maintain character consistency

Tips:

I use a training resolution of 512x512 because increasing the resolution can compromise character consistency. However, 512x512 is too small to train fine details, so it's a good idea to include close-up videos of the specific body parts you want to train in your dataset. However, if the close-up is too extreme, the model won't be able to determine which body part it is, and training will fail.