pablo-reyes8/conditioning-stable-diffusion
Latent diffusion model for attribute-conditioned face generation, with reproducible data ingestion, YAML-configured training and inference, offline evaluation through FID, KID, Inception Score, and pre-trained face detection.
What's novel
Latent diffusion model for attribute-conditioned face generation, with reproducible data ingestion, YAML-configured training and inference, offline evaluation through FID, KID, Inception Score, and pre-trained face detection.
Code Analysis
17 files read · 5 roundsA from-scratch latent diffusion training and inference pipeline that generates faces conditioned on binary attributes (e.g., Smiling, Young, Male) using a custom UNet with cross-attention, Classifier-Free Guidance, and a pretrained Stable Diffusion VAE for latent encoding/decoding.
Strengths
Genuinely implements the core diffusion components (UNet, DDPM/DDIM, attention, Min-SNR loss, CFG) from scratch with careful attention to training stability (zero-init, EMA health checks, OOM recovery, gradient accumulation), and provides a complete reproducible workflow from data ingestion through evaluation with FID/KID/IS metrics.
Weaknesses
The label conditioning is limited to 11 binary attributes via a simple embedding interpolation rather than text-based conditioning, the test suite is thin (mostly smoke tests with tiny dummy models rather than verifying numerical correctness of diffusion math), and some Colab-specific code (Drive mounting) is mixed into the core training orchestrator.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
11
Contributors
1
Files
134
Active weeks
2
Repository
Language
Python
Stars
1
Forks
0
License
—