problem of mass erasing celebrities

[1wttttw1]

I really appreciate your work and have the following questions that I hope you can answer: 1. Is there quantitative tabular data for the concept of erasing celebrities？ 2. Does mass erasure apply equally to the concept of erasing objects, since I only look erase celebrities and styles primarily in paper?Thanks！

[Shilin-LU]

Thank you for your interest. The table data for celebrity concept erasure has been organized below:

Metric	1-celebrity	5-celebrity	10-celebrity	100-celebrity
ACCe_uce	0.0000	0.0000	0.0042	0.0012
ACCe_fmn	0.1134	0.8719	0.9080	0.9223
ACCe_ac	0.9095	0.8058	0.8238	0.8913
ACCe_esdu	0.0000	0.0171	0.0458	0.0406
ACCe_esdx	0.0155	0.0683	0.1649	0.2784
ACCe_sld	0.0082	0.5292	0.6337	0.8706
ACCe_sd14	0.9143	0.9673	0.9798	0.9648
ACCe_sd20	0.9426	0.9677	0.9388	0.9324
ACCe_mace	0.0040	0.0123	0.0285	0.0430

Metric	1-celebrity	5-celebrity	10-celebrity	100-celebrity
ACCs_uce	0.9413	0.9083	0.8351	0.3790
ACCs_fmn	0.7010	0.9099	0.9091	0.9076
ACCs_ac	0.9213	0.8852	0.8692	0.8905
ACCs_esdu	0.4397	0.0970	0.1584	0.0309
ACCs_esdx	0.8947	0.5773	0.5252	0.2793
ACCs_sld	0.7920	0.7749	0.7644	0.7946
ACCs_sd14	0.9388	0.9388	0.9388	0.9388
ACCs_sd20	0.9233	0.9233	0.9233	0.9233
ACCs_mace	0.9169	0.9173	0.9248	0.8456

Metric	1-celebrity	5-celebrity	10-celebrity	100-celebrity
Hc_uce	0.9698	0.9519	0.9084	0.5495
Hc_fmn	0.7830	0.2246	0.1671	0.1431
Hc_ac	0.1648	0.3185	0.2930	0.1937
Hc_esdu	0.6108	0.1766	0.2718	0.0598
Hc_esdx	0.9375	0.7129	0.6449	0.4027
Hc_sld	0.8807	0.5857	0.4953	0.2226
Hc_sd14	0.1571	0.0632	0.0395	0.0679
Hc_sd20	0.1081	0.0624	0.1148	0.1260
Hc_mace	0.9548	0.9512	0.9476	0.8979

Metric	1-celebrity	5-celebrity	10-celebrity	100-celebrity
FID_uce	13.5177	12.7351	24.7264	106.5688
FID_fmn	13.6419	13.9330	13.9340	13.9503
FID_ac	14.1468	14.0898	13.9782	13.9181
FID_esdu	15.8395	15.0515	15.7929	15.1353
FID_esdx	14.4579	14.2121	14.2611	14.6528
FID_sld	16.3881	17.3339	17.3304	17.5449
FID_sd14	14.0432	14.0432	14.0432	14.0432
FID_sd20	14.8702	14.8702	14.8702	14.8702
FID_mace	13.2838	14.1942	14.1222	12.8215

Metric	1-celebrity	5-celebrity	10-celebrity	100-celebrity
CLIP_uce	31.1635	30.7172	29.1864	19.1703
CLIP_fmn	31.1861	31.3118	31.3088	31.3129
CLIP_ac	31.3601	31.2601	31.2610	31.2329
CLIP_esdu	30.0941	29.3750	29.1833	29.0237
CLIP_esdx	30.6987	30.3136	30.2850	29.8995
CLIP_sld	30.8538	30.9933	30.9737	30.9345
CLIP_sd14	31.3443	31.3443	31.3443	31.3443
CLIP_sd20	31.5326	31.5326	31.5326	31.5326
CLIP_mace	31.0761	30.9399	31.0343	30.2126

Metric	1-celebrity	5-celebrity	10-celebrity	100-celebrity
Ratio_uce	0.9760	0.9400	0.9800	0.7192
Ratio_fmn	0.9520	0.9680	1.0000	0.9940
Ratio_ac	0.9720	0.9680	0.9760	0.9940
Ratio_esdu	0.5760	0.4680	0.5240	0.4724
Ratio_esdx	0.7760	0.6440	0.7520	0.8088
Ratio_sld	0.9800	0.9600	0.9720	0.9796
Ratio_sd14	0.9800	0.9800	0.9880	0.9876
Ratio_sd20	0.9760	0.9920	0.9800	0.9884
Ratio_mace	0.9840	0.9720	0.9800	0.8440

[Shilin-LU]

I really appreciate your work and have the following questions that I hope you can answer: 1. Is there quantitative tabular data for the concept of erasing celebrities？ 2. Does mass erasure apply equally to the concept of erasing objects, since I only look erase celebrities and styles primarily in paper?Thanks！

To erase a large number of objects, the approach is similar to erasing other types of concepts. However, erasing a significant number of objects that form part of the prior knowledge in a pretrained diffusion model can severely damage the generative prior. For this reason, we only conducted our multi-concept erasure experiments on celebrity and style-related concepts, as these are less likely to negatively impact the generative prior.