ModKie’s vision training path uploads datasets to S3, runs YOLO training on RunPod or Ultralytics Platform, and registers the best checkpoint as an API-callable model. This closes the loop from labeled images to production inference without leaving the platform billing context.
Typical steps: prepare dataset → start training job from /train → monitor GPU job → download or auto-register weights → validate with Playground → integrate via createTask if exposed through the gateway.
Charges for training differ from async generation; credits may be pre-held by GPU tier.