At the Bar
Random shots were generated locally using Tencent’s new Hunyuan Open Source Text-to-Video Model in ComfyUI, put together in a quick edit. The purpose of this demo is to showcase the new Gen AI Video model and its impressive motion fidelity and realism at this early stage. The model can handle shots and countershots in one take, keeping both the character and location consistent. However, it’s still unable to maintain consistency across multiple takes. A lot of the cuts between the man and woman were made by the model within one 5-8 second clip. In terms of rendering, a 5-second shot at 1024x576 resolution and 16fps takes approximately 8 minutes to render on an RTX4090 with 30 steps. Reducing the resolution to 540x288 can bring the render time down to around 2 minutes per shot while maintaining reasonable quality. You can also lower the steps to 20 or fewer to further reduce the render time, still yielding good quality. For enhanced resolution, additional upscaling with Topaz Video AI is recommended. Lastly, a bit of grain was added with Filmconvert Nitrate for a more cinematic finish.
