Reproduce “HyperHuman: Hyper-Realistic Human Generation with Latent Structural Diffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu and 1 more
HYPERHUMAN offers a two-stage framework for generating hyper-realistic human images with coherent anatomy and diverse layouts by simultaneously denoising RGB, depth, and surface-normal maps in a latent structural diffusion model. This approach yielded a 25.2% improvement in FID (17.18) and a 48.5% improvement in KID (4.11x1k) over the previous best method on human image generation tasks.
Key insights
- Hyper-realistic human image generation requires explicitly modeling multi-level structural information (pose, depth, surface-normal) beyond what text prompts alone can provide.
- Simultaneously denoising RGB, depth, and surface-normal maps within a unified diffusion model significantly improves anatomical coherence and image quality by enforcing joint learning of appearance and geometry.
- A two-stage pipeline, combining a latent structural diffusion model with a structure-guided refiner and robust conditioning, effectively leverages predicted structural information for high-resolution output while mitigating error propagation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.