The authors formalize physical hallucination as two separate failure modes.
Video sampling needs both model likelihood and external physics guidance
The authors separate physical hallucinations into two risks, then design a sampler and scorer that target those risks with complementary signals.
Chinese Tech
Chenyan Zheng · Fan Wu · Ruibo Li · Jiaxing Li · Minghao Yang · Bang Zhang · +1 more
Zhejiang University · Nanyang Technological University · Tsinghua University · Tongyi Lab, Alibaba Group
Research Digest··3 min read
Zheng et al.
§