xie-lab-ml/Golden-Noise-for-Diffusion-Models

Introduces a learnable noise optimization framework that generates task-specific "golden noise" for diffusion models through paired noise prediction networks (supporting architectures like SVD-UNet+UNet and DiT). Trains on prompt-conditioned noise pairs collected via reward models (PickScore, HPS v2, ImageReward) to improve generation quality across SDXL, SD2.1, and other pipelines. Provides end-to-end utilities for data collection, multi-GPU training with gradient accumulation, and inference with classifier-free guidance integration.

194 stars.