NVIDIA Cosmos3 Nano generates 720p video in 9 minutes on dual RTX 3090s

A Reddit user benchmarked NVIDIA's Cosmos3 Nano model using vllm-omni on a system with dual RTX 3090s (48GB VRAM total), a Ryzen 9 9950X, and 128GB DDR5 RAM. With tensor-parallel-size 2 and SAGE_ATTN diffusion attention backend, the setup generated a 720x720, 161-frame video in exactly 9 minutes using 20 inference steps and guidance scale 6.0. The prompt requested a stop-motion yarn turtle video; the output was saved locally. However, attempting 1280x720 resolution caused an out-of-memory error during video decoding, though the generation phase completed without issues.

Quality-wise, the Nano variant underperforms for text-to-video (T2V) tasks, exhibiting noticeable smearing and object distortion. Image-to-video (I2V) results are ‘totally acceptable’ by the tester's account, though artifacts remain. The user highlighted the model's potential for local video generation but noted that professional-grade results may require the larger Cosmos3 model. The entire stack used vllm-omni's omnimodel support and no guardrails config. This DIY setup demonstrates that local video generation is feasible with consumer hardware, but resolution and quality trade-offs are still significant.