where can we download the weights of MIMR reward?

[JoshonSmith]

1. where can we download the weights of MIMR reward?

# detect fact to get bbox
self.det_net = init_detection_model('retinaface_resnet50', half=False, device=device, model_rootpath='models/pulid/facexlib').requires_grad_(False).to(device, dtype=weight_dtype)
# get face emb
self.arcface_model = get_arcface_model(model_name='r50', pretrained_path='models/pulid/glint_cosface_res50.pth').to(device, dtype=weight_dtype)

2. How can we use MIMR class if we have two or more reference human images?

def score_grad(self, image: torch.Tensor, image_gt: torch.Tensor, **kwargs):

[cb1cyf]

1. det_net would be downloaded automatically with function init_detection_model; you can download arcface_model from https://github.com/deepinsight/insightface/tree/master/recognition/arcface_torch

2. image is the tensor of generated image, image_gt is the tensor of ground truth image which contains all reference identities. If you want to use MIMR without the ground truth image, i.e., compute MIMR between generated image and several reference images, a simple way is getting face embeddings for each reference image, then concat or stack them to get face_emb_gt in https://github.com/bytedance/UMO/blob/main/reward_models/MIMR/multi_id_matching_reward.py#L88