Sampling Process Diagnostics¶
Current validation sampling uses a cosine-guided noise schedule.
Logged diagnostics include:
- intermediate denoising frames
- MAE vs reverse denoising step (using per-step x0 prediction)
- diffusion schedule profiles (sqrt(alpha_bar_t), sqrt(1-alpha_bar_t), beta_tilde_t, log10(SNR+eps))
Observed DDPM tradeoff: - many early steps remain highly noisy - compute is spent on low-visual-information stages
Potential improvement directions:
- DDIM sampling for faster useful denoising trajectory
- alternate schedules
- parameterization choices (x0 vs epsilon)
Intermediate reconstructions over the denoising path:
MAE trend across intermediate denoising steps:
Implemented schedule options in code:
- linear
- cosine
- quadratic
- sigmoid
Example schedule profile image:
