Why Locking the Seed Still Won't Get You the Same AI Video Shot Twice

A seed in AI video generation fixes only the starting noise pattern — not the model weights, sampler behavior, or GPU math that shape the final output. Frank Houbre, founder of Outerframe Studio and creator of ScreenWeaver, illustrates this with a concrete test: running the same prompt, seed, and character reference on his production Lost Garden four days apart returned a subtly different jawline and an unexpected camera drift. The culprit was a silent model update between runs, not any change in his settings. Three systemic causes break reproducibility even with locked seeds: model weight updates shipped between named versions, ancestral samplers that inject fresh randomness at every denoising step rather than just the first, and floating-point precision differences across GPU batches.

Houbre notes that even Runway's own documentation concedes seeds deliver "highly similar" results rather than pixel-perfect reproduction, and flags version drift and cross-platform incompatibility as top beginner mistakes. His practical conclusion — drawn from months of building the project — is that reference-based generation is a more reliable strategy for maintaining shot continuity in long-form AI production than relying on seed archives. The piece is a useful corrective for any filmmaker or post-production team building reproducibility assumptions into an AI video pipeline.