Model

DepthDif's maintained model is PixelDiffusionConditional, a conditional pixel-space diffusion model with a ConvNeXt-style U-Net denoiser.

Model-facing fields

The active dataset returns 128×128 patches. A scalar scenario has 50 generated depth channels and 53 condition channels; the joint scenario concatenates temperature and salinity into 100 generated and 103 condition channels.

Dense surface conditioning uses one scenario-derived channel: OSTIA sea-surface temperature (sst) for temperature and joint scenarios, or sea-surface salinity (sss) for the salinity scenario.

Sparse depth-aligned EN4/ARGO values, observation masks, ocean/bathymetry support, coordinates, and date context complete the condition. The scenario resolver owns the final channel contract.

Conditioning and denoising

Coordinates are encoded with the configured scheme and the date is represented by periodic day-of-year features. The resulting context is injected through FiLM in ConvNeXt blocks throughout the denoiser. See Coordinate and date conditioning.

For diffusion timestep t, the forward process is

\[ x_t = \sqrt{\bar\alpha_t}x_0 + \sqrt{1-\bar\alpha_t}\epsilon, \qquad \epsilon \sim \mathcal{N}(0,I). \]

The configured objective determines whether the model predicts noise or the clean target. Current pixel presets use the values recorded in their committed YAML snapshots; checkpoints must be loaded with a compatible objective and channel layout.

Training objectives

• Direct paired supervision uses the dense GLORYS target.
• Synthetic pretraining uses observed surfaces plus a fitted monthly/spatial GLORYS depth-delta prior, with exact ARGO anchors.
• Ambient training further corrupts sparse observations for the condition while computing loss on the original observed support intersected with valid target and ocean support.

Coastal distance weighting exists but is disabled in all current pixel presets. Optional observation, increment, structure, and spectral losses are also disabled by default. Feature Gram configuration is reserved but not implemented.

Sampling and output

DDPM and DDIM samplers are implemented. Training validation uses 100-step DDIM; the inference super-config defaults to DDPM unless explicitly overridden.

Prediction applies the target field's inverse normalization, preserves invalid support as NaN internally, and lets exporters apply final land/nodata masks. Repository exporters can retain all native depth channels or select requested physical depths by nearest GLORYS channel.

EMA weights are maintained by current pixel presets and can be compared with standard weights during validation. See EMA, Losses, and Uncertainty.