Research Context

ParTY is a CVPR 2026 text-to-motion paper on part-guided motion generation, where the model needs to reflect body-part-specific prompts while keeping the full-body motion coherent.

• The qualitative figures needed to show whether each model actually followed part-level instructions in the generated human motion.
• This made visualization quality important: unclear renders would make it harder to compare body-part behavior, motion direction, and failure cases.
• I focused on the qualitative visualization side of the paper, especially how generated motions were rendered, compared, and prepared for figures.

My Contribution

I worked on the motion visualization and qualitative comparison process used in the paper.

• Rendered the human motion outputs used in the paper’s qualitative figures, including comparisons across HumanML3D, MoMask, ParCO, and ParTY settings.
• Improved lower-quality baseline renders into figure-ready outputs through clearer body-part coloring, controlled camera setup, SSAO, anti-aliasing, and high-resolution rendering.
• Compared model outputs under matched prompts and seeds, then prepared visual materials and crops so model differences and failure cases could be shown clearly in the paper.

Visualization Comparison

Before — baseline motion visualization

After — improved motion visualization

Output

The work contributed to the qualitative comparison materials used in the final ParTY paper.

• Paper: ParTY: Part-Guidance for Expressive Text-to-Motion Synthesis.
• Venue: CVPR 2026.