EviDiff: Efficient Multi-Object Composition with Evidential Supervision
The trained EviDiff models are available at https://huggingface.co/EviDiff
1. Additional qualitative examples
2. Hyperparameters analysis
$\lambda_1$ OA↑
FID(C)↑
FID(F)↑
CLIP(C)↑
FID(F)↑
5e-4
0.7109
20.11
55.26
0.3229
0.3330
5e-5
0.7116
20.09
55.12
0.3241
0.3348
5e-6
0.7110
20.13
55.24
0.3237
0.3336
$\lambda_2$ OA↑
FID(C)↑
FID(F)↑
CLIP(C)↑
FID(F)↑
5e-6
0.7105
20.17
55.33
0.3230
0.3328
5e-7
0.7116
20.09
55.12
0.3241
0.3348
5e-8
0.7097
20.14
55.15
0.3234
0.3343
$\lambda_3$ OA↑
FID(C)↑
FID(F)↑
CLIP(C)↑
FID(F)↑
5e-6
0.7114
20.17
55.33
0.3230
0.3328
5e-7
0.7116
20.09
55.12
0.3241
0.3348
5e-8
0.7097
20.14
55.15
0.3234
0.3343
$\eta_1$ OA↑
FID(C)↑
FID(F)↑
CLIP(C)↑
FID(F)↑
1e-3
0.7109
20.31
55.54
0.3226
0.3324
1e-4
0.7116
20.09
55.12
0.3241
0.3348
1e-5
0.7102
20.28
55.32
0.3230
0.3315
$\eta_2$ OA↑
FID(C)↑
FID(F)↑
CLIP(C)↑
FID(F)↑
1e-3
0.7103
20.46
55.44
0.3213
0.3320
1e-4
0.7116
20.09
55.12
0.3241
0.3348
1e-5
0.7088
20.38
55.65
0.3219
0.3314
3. Visualization of ablation study
