Human-aware Design

Accepted at ACM Multimedia 2026

Human-aware
Design Generation.

Adding 3D Humans into Graphic Designs

Zijin HouYing Cao

ShanghaiTech University

Paper / arXiv GitHub Hugging Face BibTeX
An orange partial poster becomes a complete design: the method predicts a 3D pose, framing, and layout, places a posed 3D human into the poster, and renders a realistic human image.
The idea A human pose is part of the composition. Given a partial design, predict the human’s pose, framing, and placement together.

01 / Overview

A human that belongs
in the design.

Pose, framing, and placement shape how we read a graphic design. We make them part of the generation process.

Human images in graphic designs are carefully composed to guide attention and convey a feeling. We introduce design-conditioned 3D human adding: given a partial design with images, text, and vector shapes, generate a 3D human representation that works with the existing elements.

The model jointly predicts the human’s 3D pose, 2D framing, and image layout. The resulting representation can be integrated into the design and rendered as a realistic human image.

Explore the full paper

02 / Method

From elements to a complete design.

Pose · Framing · Layout
Model architecture: an encoder embeds the partial design’s elements. A layout generator predicts human-image bounding boxes and z-order. An autoregressive pose-and-framing generator predicts pose tokens and framing tokens, and a pretrained pose decoder reconstructs 3D humans.
Model overview Element-wise design embeddings condition a pose-and-framing generator and a layout generator.
1

Understand the design

Encode the multimodal elements of a partial design into element-wise representations.

2

Compose the human

Generate pose and framing tokens, alongside the human image’s bounding box and z-order.

3

Bring it together

Decode the 3D pose, integrate it into the composition, and render the human image.

03 / Visual results

Designed in context.

Figures from the paper.
Select a figure to view full resolution.

Qualitative comparison of 3D pose, 2D framing, and human-image layout across different methods.

Paper comparison showing partial input designs and the human poses and image layouts predicted by different methods.
Qualitative comparison Poses, framing, and placement in the context of the existing design.

Code release

Try the pipeline.

The repository includes deployment instructions and a small example bundle. The current software release extends the pipeline with a frozen FlexDM design encoder, a Transformer layout encoder, and optional photorealistic rendering.

Qwen Image 2.1 is the default realistic-rendering backend, using the white model for pose and the input design for style. These release extensions are documented separately from the paper method; the figures on this page are the original paper figures.

Installation & examples

04 / Citation

Cite this work.

DOI: 10.1145/3767308.3836019

If you find this work useful, please consider citing our paper.

BibTeX
@inproceedings{hou2026humanaware,
  title     = {Human-aware Design Generation: Adding 3D Humans into Graphic Designs},
  author    = {Hou, Zijin and Cao, Ying},
  booktitle = {Proceedings of the 34th ACM International Conference on Multimedia},
  year      = {2026},
  doi       = {10.1145/3767308.3836019}
}