Null Conditioning Signal Masking

via “null-conditioning signal masking”

* ⭐ 08/2022: [Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation (DreamBooth)](https://arxiv.org/abs/2208.12242)

Unique: Uses simple random masking of conditioning signals during training (replacing with null tokens) rather than separate data streams or model branches, enabling efficient joint conditional-unconditional training within a single model

vs others: Simpler and more parameter-efficient than separate conditional and unconditional models, but requires careful null token design and dropout probability tuning vs. dedicated single-purpose models