Reproduce “TrioPose: Native Triple-Stream Diffusion Transformers for Pose-Guided Text-to-Image Generation” on OpenResearch

About this paper

Dian Gu, Zhengyi Yang

TrioPose introduces a native triple-stream diffusion transformer framework for pose-guided text-to-image generation that effectively addresses limb distortions and feature crosstalk in multi-person scenarios. It achieves state-of-the-art pose accuracy, with a 30% AP improvement on Human-Art, while simultaneously maintaining high visual fidelity and strong semantic alignment.

Key insights

• Treating pose as a native, independent modality within a DiT architecture, combined with zero-initialized dual-residual injection, provides stable and high-fidelity pose control without disrupting pre-trained representations.
• Fine-grained topological classification and learnable continuous attention biases (LRBM) are critical for robustly decoupling inter-instance feature entanglement in complex multi-person scenes, outperforming hard binary masks.
• Targeted anatomical supervision via spatial loss weighting, activated with a time-step truncation constraint (PSLW), effectively corrects micro-anatomical distortions while preserving global layout stability and visual fidelity.

Summary by alphaXiv.