Tiny models from tiny data: Textual and null-text inversion for few-shot distillation