Cosmos3 Nano testing with vllm-omni
A user shared their experience testing the Cosmos3 Nano model for video generation using vllm-omni. They reported impressive speed, generating a 720x720 video in under 10 minutes on their dual RTX 3090 setup. However, they encountered out-of-memory errors when attempting higher resolutions during video decoding and noted that while image-to-video quality was acceptable, the nano model produced some artifacts and smearing. AI
IMPACT Demonstrates performance characteristics and limitations of open-source video generation models, informing potential users.