marcemq/crowdmod-ddpm-4D
Crowd modeling with Denoising Diffusion Probabilisitic Models (DDPMs) for crowd macroproperties sequence samples.
What's novel
Crowd modeling with Denoising Diffusion Probabilisitic Models (DDPMs) for crowd macroproperties sequence samples.
Code Analysis
12 files read · 4 roundsGenerates and evaluates crowd macroproperty sequences (density, velocity) using three generative model families—DDPM, Flow Matching, and ConvRNN—with UNet and DiT backbones, domain-specific guidance, and comprehensive motion-based metrics.
Strengths
Genuinely novel DiT4D_V4 architecture with factorized spatial-temporal attention and AdaLN-Zero modulation; complete research pipeline with multiple model families, proper diffusion/flow-matching math, and thorough domain-specific evaluation metrics.
Weaknesses
No test suite at all; a few implementation bugs (Heun integrator uses delta_k2=1 instead of delta, mass-preservation gradient computed via an O(N) Python loop over all spatial elements).
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
48
Contributors
3
Files
76
Active weeks
23
Repository
Language
Python
Stars
1
Forks
1
License
CC0-1.0