ByteDance Proposes Generative Refinement Networks as a Third Route for Visual Generation

挑战扩散自回归统治！字节提出视觉生成第三种路线，让模型像人类一样边画边改

ByteDance’s commercial technology team proposed GRN, a visual generation architecture using HBQ, global refinement, and complexity-aware sampling to address quantization loss, error accumulation, and fixed-step inference; on a 130M model, adaptive sampling reduced inference from 50 steps to an average of 24, while gFID changed from 3.56 to 3.79.

Why it matters: HKR-H/K/R all pass: ByteDance’s GRN has a concrete hook plus 130M, 24-step inference and gFID 3.79. It is a strong research release, not a flagship model launch, so it stays in the 78–84 band.