OpenAI historically does not publish exact training or alignment recipes for frontier models; Sora follows the same pattern.

Claim “Sora uses DPO” → Unsupported

Claim “Sora does not use DPO” → Also unsupported

Only defensible position: Unknown / not publicly specified

So under strict market or resolution rules:

👉 N/A is the correct resolution if the criterion is verifiable public confirmation by EOY 2025.

DPO can be extended to diffusion models by contrastively optimizing the probability of Gaussian noise. 

By Sora, I am referring to the version used in the 2024 Feb publication. 

Resolve NA if I cannot figure out by EOY 2025

People are also trading

Related questions