1. Keep the workflow reproducible
Record ComfyUI commit, custom-node commits, model checksums, resolution, sampler, step count, seed and batch size. Keep models, inputs, outputs and workflow JSON in persistent storage with a separate backup.
2. Start with a small SDXL batch
A 24GB GPU is a useful starting point, but ControlNet, upscalers, video nodes and large batches can change memory demand. Warm up once, then measure at least several representative batches and record failures.
nvidia-smi --query-gpu=timestamp,memory.used,utilization.gpu --format=csv -l 23. Calculate cost per successful image
GPU time cost = billed hours × hourly rate. Total cost = GPU time + storage + transfer + other fees. Divide by successful outputs, not requested images. Example: 2 hours at $1.20/hour is $2.40; 240 successful outputs gives $0.01/image before other charges. These are illustrative inputs, not a performance claim.
4. Secure access and finish the job
Bind the UI to localhost or a private interface and use an authenticated TLS gateway. Export workflow JSON, download outputs and verify backups before stopping or deleting the instance. Confirm whether stopped GPUs, attached disks or snapshots continue billing.