Lora from HunyuanVideo is currently the open-source model that can make characters 99.99% similar to those in the show. I want to say it can instantly boost your training level by at least one grade, like from A to A+, or even directly to A++. I'm really curious why the previous Hunyuan image models couldn't achieve this.
Training Lora for HyV is easy, but the testing phase is really painful. We need much more powerful GPUs!