Why AI Characters Drift Between Shots

You generate a character, they look perfect in shot one, and by shot four they're someone else's slightly off cousin. Why? This episode digs into the structural reasons image-to-video models drift: stateless sampling, category-level text prompts, and the fundamental tension between identity consistency and scene diversity. We cover when training a LoRA actually makes sense (and when it's a day of setup for nothing), what the ideal single reference image looks like, why matching your reference background to your scene backfires, and the counterintuitive truth about multi-reference conditioning — more angles can make drift worse. If you've ever wondered why the model "fails" at character consistency, this reframes the whole problem. Episode #121607 — open it directly at myweirdprompts.com/121607