3D_Brain_partsLDM_DDIM50_PURE_NOISE

3D Brain partsLDM — Aux-tAux with a 50-step pure-noise DDIM scaffold

Checkpoints for the DDIM50_PURE_NOISE ablation of the whole-brain compositional latent diffusion model (spacing 1.5 mm, 128^3).

What is different in this variant

The Aux-tAux branch needs a whole-brain scaffold latent at an auxiliary timestep t_aux. The baseline builds it cheaply with forward_noise, i.e. q(z_0, t_aux) applied to the encoded ground-truth volume. This variant instead builds the scaffold from pure Gaussian noise: starting at z ~ N(0, I) at t = 999, it runs 50 deterministic DDIM transitions (eta = 0) down to the sampled raw t_aux, using the online whole-brain UNet with gradients detached. The shared noise used to map the composite back to t_aux is the model's epsilon prediction at the landing point, not a forward-process noise tensor.

Motivation: the scaffold the auxiliary loss sees at training time then comes from the same distribution the model actually rolls out at inference, instead of from a ground-truth-anchored forward diffusion.

Training setup

Initialized from the spacing-1.5 whole-brain UNet at step 15000, the whole-brain AE, and the frozen hemisphere / subcortical part experts (April AE checkpoints + their step-15000 UNets).

Contents

• ckpt_whole_brain_aux_taux_UNET_spacing1p5_DDIM50_PURE_NOISE.tar — final weights (UNET_step000002500.pt, UNET_last.pt) plus the full args.json of the run.
• UNET_loss.png — training loss curves.
• eval_samples/step000000625|001250|001875|002500/ — in-training samples at each eval checkpoint (2 samples per round).

Untar with:

tar -xf ckpt_whole_brain_aux_taux_UNET_spacing1p5_DDIM50_PURE_NOISE.tar

Note

Research artifact from an ablation study; not validated for any clinical use.