Sampling¶

Training always uses the full configured forward schedule. Sampling chooses how to traverse that learned clock in reverse.

DDPM¶

DDPM_Sampler visits every training timestep in descending order and samples from the exact posterior. It is ancestral and stochastic except at the final step.

model:
  num_timesteps: 1000
  sampling:
    method: ddpm
    # timesteps must equal num_timesteps

DDPM is the implicit default when model.sampling is absent. It is faithful to the full chain but expensive during validation.

DDIM¶

DDIM_Sampler builds an endpoint-inclusive, evenly spaced subset of actual training timesteps and stores each scheduled predecessor. The model and sampler therefore agree on timestep embeddings and coefficients.

model:
  num_timesteps: 1000
  sampling:
    method: ddim
    timesteps: 100
    eta: 0.0
    clip_sample: true

With eta: 0, the chain is deterministic given initial noise. Positive eta adds controlled stochasticity. DDIM allows fewer steps, making reconstruction logging much cheaper.

Ensembles and classifier-free guidance¶

Different initial Gaussian latents produce different conditional samples even when DDIM uses eta: 0. validation.ensemble_size assigns each sample a stable latent namespace, making diversity and calibration comparable across epochs. The model exposes sample_ensemble(...), which returns tensors shaped [member, batch, channel, height, width].

When condition dropout was enabled during training, sampling.guidance_scale combines conditional and zero-condition noise predictions. A value of 1 performs the original single conditional evaluation. Values above 1 usually increase fidelity to the baseline and GEO/ERA context at some cost to ensemble diversity.

Clean-sample clipping¶

Both samplers estimate the clean target at every step. When clip_sample: true, that estimate is clamped to [-1,1] before calculating the previous sample. This limits runaway errors at low signal-to-noise timesteps and matches the model’s training range.

Reproducibility¶

Diffusion validation noise is stable per sample, not merely per batch order. A SHA-256 digest of the validation seed and sample ID seeds a same-device torch.Generator. Consequently:

• the same sample receives the same initial latent across epochs;
• storm-stratified ordering does not change the latent;
• global RNG state is not disturbed; and
• DDIM with eta 0 gives directly comparable reconstructions.

DDPM still draws reverse noise. Its sampler accepts an explicit generator, but the current _predict_batch() supplies fixed initial noise rather than a reverse-process generator; complete DDPM chains can therefore remain stochastic.

Schedules¶

Changing a schedule changes the learned problem. Start a fresh run or resume with the checkpoint’s original schedule; code explicitly rejects coefficient mismatches.