Cheng et al.
Combining compression and adaptive inference for diffusion transformers without quality loss
TamedFlow trains a compressed Diffusion Transformer for multi-step drafting, using a lightweight router to decide when to reuse student predictions or revert to teacher help.
Chinese Tech
Shuaiqi Cheng · Jiahao Zhang · Boyuan Gu · Siyu You · Yuxi Chen · Haiying Xu · +3 more
The Hong Kong University of Science and Technology (Guangzhou) · University of Electronic Science and Technology of China · The Hong Kong University of Science and Technology · Huawei Technologies Ltd. · Guangxi Zhuang Autonomous Region Information Center
Research Digest··3 min read
18 on ImageNet 256×256).
§