DiT

Embedding Prediction Helps Image Generation: NEPA-DiT Achieves Superior FID with Much Less Compute

預測嵌入向量助力影像生成：全新 NEPA-DiT 框架以超低算力達到更優 FID

This study introduces NEPA, a framework that dynamically predicts and updates embedding conditions during denoising, enabling NEPA-DiT-XL to achieve an impressive 1.32 FID with only one-third of REPA's training compute.