LoKr model trained carefully on two subjects:

• Cheech Marin

• Tommy Chong

The dataset mostly contains images of the pair together, captioned with the positional information, eg. Cheech Marin sitting to the left of Tommy Chong.

Despite my best efforts at captioning, I think Flux's internal bias here is something to work with instead of working against. For example, asking for Tommy Chong sitting to the right of Cheech Marin doesn't work as well as Cheech Marin sitting to the left of Tommy Chong.

Similarly, the idea Flux has for positional information seems to be first-person possessive and not third-person. The 'left' and 'right' directions are with respect to the character in the scene. In other words 'on the right of tommy chong' looks to us as if it's on the left. Backwards fun time!

(20s/early 30s) "young", (more recent) "elderly" variants are included when you modify your prompt.

90s sitcom trigger word should make Chong look more like the That 70's Show era.