search

joeyz0z / ConZIC

Official implementation of "ConZIC: Controllable Zero-shot Image Captioning by Sampling-Based Polishing"
MIT License
71 stars 17 forks source link

about prompt #3

Open Rouchestand opened 1 year ago

Rouchestand commented 1 year ago

Dear author, I want to achieve the description effect shown in Figure 6, that is, generate sentiment control text without prompt "Image of a". "I set the super parameter to null characters, but the effect is not good. Could you give me some guidance? Thank you very much.

Rouchestand commented 1 year ago

33658c52-7d46-433b-9860-eaef7d4e5d83

joeyz0z commented 1 year ago

For prompt: actually, we just omitted the fixed prefix “Image of” for better presentation in illustrations. In experiments, we empirically find that a good prompt like "Image of a" will significantly improve the quality of generations. For sentiment control: for more obvious effects, you can appropriately increase the controllable hyper-parameter gamma. But not too much, or you will end up with a lot of repetitive sentiment words like "perfect wonderful perfect".

Rouchestand commented 1 year ago

Thank you for your enthusiastic answer.