SMPL-GPTexture: Scalable, Training-Free SMPL Texture Synthesis With World Knowledge Transfer From GPT via Geometry-Aware Projection

Abstract

Texturing 3-D human avatars from natural-language prompts remains challenging because generated UV texture maps must be both high-quality and semantically consistent with the prompt. Prior work typically retrains generative adversarial or diffusion models on curated datasets, which constrain generalization to clothing, stylistic, and cultural variations present in those datasets. To address these limitations, we present SMPL-GPTexture, a training-free pipeline that leverages the rich world knowledge of a pretrained GPT model to synthesize SMPL human textures. Given a tailored natural-language prompt, GPT-4o generates geometry-consistent front/back views; we fit SMPL to each view, and invert-rasterize colors into canonical UV space with a geometry-aware view-quality weight, and close remaining gaps with diffusion-based inpainting. On a 550-prompt benchmark, SMPL-GPTexture achieves +17.6% CLIP similarity and +22.4% BLIP improvement over state-of-the-art text-to-texture methods while reducing visible seam artifacts. It generalizes to diverse scenarios such as fantasy avatars, stylized body art, and specialized uniforms without domain-specific fine-tuning. This facilitates the generation of high-quality textures that surpass current text-to-texture pipelines for fashion prototyping, AR telepresence, and digital asset generation. We open-source a UV texture map dataset that covers 23 artistic styles, 49 professions, 19 ethnicities, and 67 clothing descriptors, and a set of easily adaptable prompt templates that can be repurposed to generate any desired texture, fostering research on prompt-driven avatar appearance.