State of open video generation models in Diffusers

Hugging Face · 1 year ago · 44

Open-source video generation models including CogVideoX, Mochi-1, and LTX Video have proliferated alongside proprietary offerings from OpenAI, Google, and others, prompting Hugging Face's Diffusers library to provide optimization tools for their deployment. HunyuanVideo requires 60.09 GB of memory under standard settings but can be reduced to 6.56 GB through combined quantization, CPU offloading, and tiling techniques. Users can now run video generation models on consumer hardware by chaining multiple optimization strategies, though doing so extends inference time from 863 seconds to around 1000 seconds.