search

PeterWang512 / GenDataAttribution

Evaluating Data Attribution for Text-to-Image Models: a visual data attribution benchmark for evaluating and learning training image influences.
https://peterwang512.github.io/GenDataAttribution
Other
58 stars 3 forks source link

Questions about the mapper weight #2

Open kaoshizhou opened 10 months ago

kaoshizhou commented 10 months ago

Hi author,

Thanks for your wonderful work!

I have downloaded the mapper weights from here, and I found that there is only one weight file for each image encoder. I would like to sincerely ask, does this weight file represent the "Object+Style" model in the paper?

If so, would you release the "Object-Centric" and "Artist-Style" models in the near future? I'm very interested in performance comparisons between models fine-tuned on different sets (e.g., DINO Object-Centric outperforms DINO Object + Style on style Recall@10 as shown in Figure 5 in the paper, which is a bit counterintuitive).

Thanks again for your contribution to the community! Looking forward to your reply.

PeterWang512 commented 9 months ago

Hi,

Sorry for the late response. Yes, the weight files are all "Object+Style" models. I also updated the repo so that the "Object-Centric" and "Artist-Style" models can be downloaded by running bash weights/download_style_object_ablation.sh. You might need to git pull the latest version of the repo. Thanks!