Post-caption trick: train LoRA with activation word only, caption at inference

Inevitable_Pen9043 · reddit · 2026-11-20

The author shares an experiment: 1) train a LoRA using only an activation word with no captions; 2) caption one dataset image with JoyCaption; 3) apply that caption at inference. The surprising hint: even without captions in training, the model may still learn the concepts, and you can steer them at inference time.

The author isn't fully sure of the mechanism, but this "post-caption" approach suggests you can skip captioning during training and refine outputs with captions only when generating.

Original post →

More from Multimodal

Multimodal channel →