Pro-Pose: Unpaired Full-Body Portrait Synthesis via Canonical UV Maps

2025-11-04 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

Pro-Pose is a novel framework developed by researchers at The University of Texas at Austin and Google for generating high-fidelity, reposed full-body portraits from a single "in-the-wild" photograph. It addresses the challenge of limited paired training data by transforming input photos into a canonical UV space, coupled with a "Donor-based UV Reposing" methodology to model occlusions and novel view synthesis. This approach leverages existing unpaired datasets (approximately 470K images) alongside scarce paired data (30K images from DeepFashion). Pro-Pose also personalizes outputs via multi-image finetuning and achieves state-of-the-art qualitative and quantitative performance on real-world imagery, including DeepFashion (8570 pairs) and WPose (2305 pairs) datasets. The system standardizes clothing to a black tank top and shorts using Gemini 2.5 Flash Image.

Key takeaway

For machine learning engineers developing avatar generation systems, Pro-Pose offers a robust solution to overcome data scarcity and identity drift. You should consider adopting its canonical UV space and Donor-based UV Reposing for effective pose-texture decoupling, enabling generalization from abundant unpaired data. Implement few-shot adaptation to significantly enhance identity preservation, especially for extreme pose changes or when high fidelity is critical.

Key insights

Pro-Pose generates high-fidelity, reposed human avatars from single images by decoupling pose from texture in UV space.

Principles

• Canonical UV space decouples pose from texture.
• Donor-based reposing prevents pose leakage.
• Hybrid data training improves generalization.

Method

Pro-Pose uses a Flow Matching model with a Flux.1 [dev] transformer backbone, operating in a 16-channel latent space. It employs a dual-branch training strategy for paired and self-supervised unpaired data, using Donor-based UV Reposing and optional face crop dropout.

In practice

• Generate diverse pose portfolios from one photo.
• Standardize subject's clothing to black attire.
• Fine-tune models with few images for personalization.

Topics

• Full-Body Portrait Synthesis
• Canonical UV Maps
• Donor-based UV Reposing
• Flow Matching Models
• Diffusion Models
• Few-Shot Adaptation
• SMPL-X

Code references

• black-forest-labs/flux

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• One-Shot Novel View and Pose Human Image Synthesis via 3D Prior Guided Diffusion Model — A 3D prior-guided conditional diffusion model enables high-quality one-shot novel view and pose human image synthesis.
• HumanSplatHMR: Closing the Loop Between Human Mesh Recovery and Gaussian Splatting Avatar — Jointly optimizing human pose and avatar rendering improves 3D geometry and generalization for in-the-wild video.
• Every9D-21M: Large-Scale Real-World 9D Canonicalization of Everyday Objects — Large-scale, real-world 9D pose datasets can be built efficiently by propagating sparse manual annotations across object-centric videos.
• Large-Scale High-Quality 3D Gaussian Head Reconstruction from Multi-View Captures — HeadsUp reconstructs high-quality 3D Gaussian heads from multi-view inputs using a UV-parameterized latent representation.
• DPPE: Rethinking Camera-Based Positional Encoding for Scaling Multi-View Transformers — Decoupling rotation and translation in camera-based positional encoding prevents training stagnation in scaled multi-view Transformers for NVS.
• WildSplatter: Feed-forward 3D Gaussian Splatting with Appearance Control from Unconstrained Images — WildSplatter enables fast, pose-free 3DGS from unconstrained images by disentangling geometry and appearance via learned embeddings.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.