Running MiniMax H3 Image-to-Video on a Rented A100: Setup, First Clip, and Costs

说明与用途

小团队使用 Vast.ai 租用的 A100 GPU 运行 MiniMax H3 图像到视频模型。通过预处理图像、选择完整费用字段的报价以及批量处理，将单个 5 秒剪辑的成本降至约 0.03 美元，82 个剪辑的总成本为 3.59 美元。

问题

Small content teams need cost‑effective GPU rental for AI video generation, but setup and weight download costs can erode savings.

结果

A 5‑second clip generated in 86 s on an A100 cost ~$0.03 for GPU and disk time; 82 clips in one session cost $3.59 (~$0.04 per clip when including setup).

工作流

• Prepare source image on exact output canvas and check aspect ratio
• Assemble image and prompt queue before renting
• Search Vast.ai offers with filters: reliability>=0.97, num_gpus=1, disk_space>=150, inet_down_cost=0
• Create instance with onstart script to download weights
• SSH into instance, confirm weights, run first clip to validate setup
• Batch remaining clips, monitoring GPU usage
• After completion, copy outputs, verify file count and size, then destroy instance

证据

We ran MiniMax H3 image-to-video on a rented Vast.ai A100 SXM4 80GB. A 5-second, 576×1024 clip took about 86 seconds to generate and cost roughly $0.03 in GPU and disk time.

来源

We built a public Docker image on GHCR with GitHub Actions. That build took about 11 minutes.

来源

We generated 82 hero clips in one session. The session cost $3.59 and took about three hours including setup.

来源