multimodalcritiquebearishExisting single-image head avatar reconstruction approaches struggle to preserve 3D consistency under unseen viewpointsComputer Vision02 Aug 2026http://arxiv.org/abs/2607.28164v1