Run consistent multi-shot AI product & character video on rented cloud GPUs
Avoid $2,000 local hardware and Python dependency conflicts. Lock subject identity across shots in ComfyUI and spin down cloud billing automatically on idle.
Why default setups fall short
Two things burn hours in open-source AI video: CUDA version mismatches when installing custom ComfyUI nodes on cloud notebooks, and character/product drift between shot 1 and shot 2. Generating a clean anchor composite image first and feeding it through a pinned ComfyUI API graph gives repeatable 16:9 clips for cents per render.
Prerequisites before starting
- A Google Colab Pro, RunPod, or Vast.ai L4 / A100 session
- 1-3 clean reference product or character images (PNG/WebP)
- Basic familiarity with running a shell bootstrap script
Step-by-step implementation
1. Pin Torch and custom node commit hashes in one bootstrap script
Never git clone main branches of 12 custom nodes blindly—one upstream update breaks xformers or attention backends. Pin exact commit SHAs and cache model weights onto persistent network storage.
git clone https://github.com/comfyanonymous/ComfyUI.git cd ComfyUI && git checkout 7fc3a91 pip install --no-cache-dir -r requirements.txt
2. Lock identity in still frames before generating motion
Never ask a text-to-video model to invent both subject identity and camera movement simultaneously. Render 4 candidate still frames first, pick the sharpest anchor frame, and drive image-to-video with restrained camera verbs ('slow dolly in, locked tripod, natural window light').
3. Attach a 20-minute GPU idle watchdog so you never forget a running pod
Run a lightweight background bash check monitoring nvidia-smi utilization; if GPU utilization stays under 3% for 20 consecutive minutes, flush outputs and terminate the pod automatically.