Character consistency isn't a seed trick: a 2-stage image pipeline that actually locks the face

Short summary

A practical 2-stage pipeline for maintaining character consistency across generated image scenes: first generate a canonical base image with text-to-image, then use an image-edit model with that base as reference for every subsequent scene. Seeds and prompt-only approaches fail because they don't carry identity forward; LoRA works but is too expensive for apps with on-demand character creation. The article also covers async job queues and serverless-GPU deployment patterns to make the pipeline production-ready.

• •2-stage pipeline: cast character once with text-to-image, then re-scene using image-edit model with base image as reference
• •Seeds drift on prompt change, prompt-only has no memory, LoRA is too heavy for on-demand apps
• •Includes async job queue pattern and serverless GPU setup for production deployment

Generated with AI, which can make mistakes.