Diffusion video reproducibility drifts across GPUs

The short version: fixed latent plus deterministic sampling is necessary, but not sufficient for cross-GPU reproducibility. Tiny floating-point and kernel-order differences can accumulate across many steps, so “same output” is a stronger claim than most stacks can support.

• –The biggest risk is not the seed; it is backend variance from attention kernels, matmul precision, and reduction order
• –On a stable stack, architecture differences usually show up first in textures, faces, edges, and other high-frequency details
• –Video diffusion is more sensitive than still-image generation because errors compound over more frames and more denoising steps
• –If you need reproducibility, lock the entire software stack and validate perceptual similarity, not bitwise equality
• –The practical question is whether drift stays within “same idea, slightly different render” or crosses into “different clip”; that depends on how numerically brittle the model is