Hi-DiT hybrid Latent-Pixel diffusion hits FID 1.06 on ImageNet, accepted at ECCV 2026

量子位 · wechat · 2026-09-26

Researchers from USTC and HiDream.ai propose Hi-DiT (Hybrid Latent-Pixel Diffusion Transformer), accepted at ECCV 2026.

A shared Transformer backbone coordinates a Latent stream (global structure in high-noise phases) and a Pixel stream (high-frequency detail in low-noise phases), with time-gated injection at τ=0.3.
A High-Frequency Pixel Predictor uses conv + PixelShuffle to hierarchically decode sub-patches, easing high-dimensional pixel regression.
Results: FID 1.06 on ImageNet 256×256 (CFG), gFID 1.26 on 512×512; ablations show pure Pixel stream gets 3.38, +Latent input 1.74, +predictor head 1.65.
Overhead is minimal: 17.7→18.1 min/epoch, 1.58→1.67 s/image, 35.52GB peak memory. Code open-sourced at github.com/HiDream-ai/Hi-DiT.

More from Multimodal