wezteoh/traj-diffusion

Sketch-Guided Trajectory Diffusion

A diffusion model for generating smooth and diverse trajectories conditioned on sparse sketch guidance.

This model explores sketch-conditioned trajectory simulation using denoising diffusion techniques. Given a coarse spatial sketch or trajectory prior, the model generates realistic trajectory samples that preserve the intended global structure while allowing stochastic local variation.

Blog post:
https://wezteoh.github.io/posts/diffusion-for-sketch-guided-trajectory-simulation/

Code base: Model - https://github.com/wezteoh/gameplay-trajectory-diffusion Sketch canvas - https://github.com/wezteoh/gameplay-trajectory-canvas

Overview

The model learns a conditional diffusion process over trajectory sequences:

• Encode partially observed trajectory guidance
• Add noise to trajectories during training
• Learn iterative denoising conditioned on sketches
• Sample plausible trajectories at inference time

Applications include:

• game AI movement simulation
• multi-agent gameplay strategy simulation
• synthetic behavior generation

---

Model Details

Inputs

• sparse trajectory sketches
• trajectory masks

Outputs

• generated trajectory sequences

Architecture

• diffusion transformer backbone adapted for spatiotemporal task
• DPM-solver / iterative DDPM-style sampling

---

Usage

python scripts/sample_trajectory_ddpm.py \
    --checkpoint ckpt_file_path \
    --num-samples 8 \
    --input-dir sketches_dir_path \
    --save-videos