PulseAugur
LIVE 08:19:48
research · [2 sources] ·
0
research

Researchers propose FD-loss to optimize visual generation in representation space

By PulseAugur Editorial · Summary by gemini-2.5-flash-lite from 2 sources

Researchers have introduced a new training objective called FD-loss, which optimizes the Fréchet Distance (FD) in representation spaces for visual generation. This method decouples the population size for FD estimation from the batch size used for gradient computation. Applying FD-loss to existing generators has shown improvements in visual quality, with one-step generators achieving a 0.72 FID on ImageNet 256x256 when using the Inception feature space. AI

Summary written by gemini-2.5-flash-lite from 2 sources. How we write summaries →

IMPACT Introduces a novel training objective that may improve visual generation quality and evaluation metrics for generative models.

RANK_REASON Academic paper introducing a new method for visual generation.

Read on arXiv cs.CV →

COVERAGE [2]

  1. arXiv cs.CV TIER_1 · Jiawei Yang, Zhengyang Geng, Xuan Ju, Yonglong Tian, Yue Wang ·

    Representation Fr\'echet Loss for Visual Generation

    arXiv:2604.28190v1 Announce Type: new Abstract: We show that Fr\'echet Distance (FD), long considered impractical as a training objective, can in fact be effectively optimized in the representation space. Our idea is simple: decouple the population size for FD estimation (e.g., 5…

  2. arXiv cs.CV TIER_1 · Yue Wang ·

    Representation Fréchet Loss for Visual Generation

    We show that Fréchet Distance (FD), long considered impractical as a training objective, can in fact be effectively optimized in the representation space. Our idea is simple: decouple the population size for FD estimation (e.g., 50k) from the batch size for gradient computation (…

RELATED ENTITIES

RELATED TOPICS