greyDonald/HumanAwareDesignGeneration

Human-aware Design Generation — model resources

Resources for Human-aware Design Generation: Adding 3D Humans into Graphic Designs, Zijin Hou and Ying Cao, ACM Multimedia 2026.

Paper · Code · Project page

Built with Qwen. The default final renderer is native RGBA Qwen Image 2.1, conditioned on a predicted mannequin for pose and the non-human input design for appearance. This is a resource bundle for the released implementation, not a claim that the optional Qwen/SDXL extensions were the original paper's architecture.

Contents

Directory	Purpose
humanaware/	Trained layout and human decoders; inference-only checkpoint with portable configuration
flexdm/	Frozen design encoder and structured-input vocabulary metadata
pose/	Pose-token VQ-VAE decoder
assets/camera/	Quantized camera centers
assets/fonts/	Rendering font and its permission notice
qwenimage21/	Native Qwen Image 2.1 transformer, text encoder, RGBA VAE, processor, scheduler and tokenizers
pose_diffusion/	Optional SDXL, IP-Adapter, image encoder, VAE and LayerDiffuse resources

manifest.json records relative paths, file sizes, SHA-256 checksums and public-upload eligibility. Model parameters are copied unchanged; the HumanAware checkpoint export omits training history and rewrites resource paths without changing tensors. No training dataset, validation dataset, test examples or inference results are stored here.

Use

Install the two Python environments described in the code README. From the code repository, download resources into the sibling ../weight directory:

python scripts/download_weights.py --weights-dir ../weight
# After separately installing your licensed SMPL-X model:
python scripts/run_demo.py --weights-dir ../weight --render-gt

The default download includes the core model and Qwen. Pass --include-sdxl for optional SDXL resources, or --revision <commit> for a fixed release. Code and project-page publication are subject to the repository owner's final review; this model upload does not itself publish the code.

Separately obtained resources

Two resources are deliberately not redistributed in this public repository:

• assets/body_models/smplx/SMPLX_NEUTRAL.npz: obtain directly from the official SMPL-X provider under its model license. Use the code's scripts/install_smplx.py to install your copy. It is required for white-model rendering and the default complete pipeline.
• pose_diffusion/controlnet/{config.json,diffusion_pytorch_model.safetensors}: the optional SDXL OpenPose ControlNet refers to the restrictive OpenPose license. Obtain it from the original provider under applicable terms if using SDXL. The default Qwen pipeline does not use ControlNet.

The owner's local resource bundle contains these files for authorized local use. Their entries and hashes remain in the manifest with public_upload: false so missing public downloads are explicit.

Licenses and attribution

This is a mixed-license resource bundle. See THIRD_PARTY_NOTICES.md, model_sources.json, and the individual license texts. In particular, Qwen Image 2.1 is licensed under the Qwen RESEARCH LICENSE AGREEMENT (September 20, 2026), for non-commercial use subject to its terms. SDXL and LayerDiffuse weights have their own OpenRAIL terms; permissive code or conversion licenses do not replace upstream model restrictions. No blanket license is granted to all bundled components.

Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.

Validation and limitations

The released core resources were checked through independent structured-input inference, white-model rendering and GT comparison on three original demonstration cases. Native Qwen RGBA rendering uses two image references, BF16, CPU offloading, 40 steps and no VAE tiling on an NVIDIA L40. Generated people can vary in pose detail, hand fidelity and clothing appearance; inspect results for your use case. GT is used for post-generation comparison only.

Citation

@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}
}