Are all experiments conducted with CFG/η=2?

[ssateq]

I understand that for demon, η=2 brings more diversed search spaces. But for baseline, best-of-N, and DOODL, shouldn't they perform better at like η=5?

[ssateq]

Like, does demon with cfg=2(some optimal number) outperform best-of-N with cfg=5(some optimal number)?

[rareone0602]

For DOODL (and Universal Guidance for Diffusion Models), we adhere strictly to the configurations described in the demo notebook’s config, except as stated in the paper.
We use cfg=2 for other experiments related to the baseline. While it is true that increasing the cfg value tends to yield better, more detailed outputs, we still observe that the overall aesthetic score plateaus around 7 for best-of-N.

We particularly want to observe the effect of tuning noise in an unconditional diffusion model. Regarding an ablation study on cfg, we opted not to include one, as the two methods rest on fundamentally different paradigms: Bayesian formulation vs the RL guidance. Exploring how these paradigms might interact or complement one another is an interesting question, but it falls beyond the scope of the work.