Establish performance on alternative flow-model backbones

Establish the performance of Flow Contrastive Preference Optimization on flow-model backbones other than Stable Diffusion 3.5 Medium.

Background

The empirical evaluation fine-tunes Stable Diffusion 3.5 Medium, so the reported results do not determine whether the method generalizes to other generative-model architectures or backbones. The authors explicitly state that performance on other backbones has not yet been established, leaving cross-backbone generalization unresolved.

References

We evaluate fine-tuning on SD3.5-M, so performance on other backbones remains to be established.

— FlowCPO: A Unified Divergence View of Preference Alignment for Flow Models  (2609.09905 - Han et al., 9 Sep 2026) in Section 5, Conclusion and Limitations