(a) Training x: tissue + bubbles 32 × 16 × 68 × 96 complex, baseband split over time m = meanT(x) r = x − m target (remainder) clip mean m (static tissue): free input of the decoder r̃ = 4r ε ~ N(0, I) t ~ sigmoid(N(0, 1)) yt = (1 − t) ε + t r̃ c time embedding Flow decoder vθ U-Net, 1.5 M skips L = ‖ vθ − (r̃ − ε) ‖² trains encoder and decoder together no bubble labels [r, m] 4 channels Eφ encoder, 0.80 M z 16 × 8 × 8 × 34 × 24 8× fewer values than the clip joined at the bottleneck (b) Sampling ε vθ Euler, 25 steps r̂ z, m
Figure 1. Autoencoder with a flow-matching decoder (diff_tis_bb). (a) A clip x of real tissue with rendered bubbles is split into its mean over time m and the remainder r = x − m; the two parts are companded. The encoder Eφ maps [r, m] to a latent grid z. The decoder vθ is a U-Net that gets the noisy remainder yt, the clip mean m, the time t, and z at its bottleneck, and gives the velocity from noise to data. Because the decoder gets m free, z has no need to hold the static tissue. The two networks are trained together with the rectified-flow loss; no bubble label is used. (b) A sample r̂ is made from noise with 25 Euler steps, with z and m held fixed.

Source: research/bubae/diffdec.py, train_diff.py, bb.py; run h100-7:~/bubae/runs/diff_tis_bb (3000 steps, batch 2).

(a) Flow decoder vθ (1.5 M parameters) Conv: 3×3×3 over (Y, X, Z) TConv: 3 taps over T TBlock: residual block with time embedding (c) 1×1×1 conv after a concatenation cconcatenate · ↓ stride 2 · ↑ nearest ×2 skip: decoder input, 4 ch skip: 32 ch skip: 64 ch 4 [yt, m] 32×16×68×96 3232 Conv ↓Z TBlock 32×16×68×48 646464 Conv ↓YXZ TConv ↓T TBlock 16×8×34×24 9696 TConv ↓T TBlock c 969696 Conv 1×1 TBlock TBlock 8×8×34×24 (latent grid) Conv 3×3×3, 16→96 z latent, 16 ch 8×8×34×24 ↑T c 646464 TConv Conv 1×1 TBlock 16×8×34×24 ↑T 64 TConv ↑YXZ 32×8×34×24 c 323232 Conv Conv 1×1 TBlock 32×16×68×48 ↑Z c 322 Conv GroupNorm · SiLU · Conv (starts at zero) 32×16×68×96 2 velocity v (b) Time embedding t sin, cos (64) Linear → 128 SiLU Linear → 128 e One vector e for each clip. Each TBlock adds it through its own linear layer. (c) TBlock, C channels h GN · SiLU Conv GN · SiLU TConv out residual e Linear 128 → C
Figure 2. The flow decoder in detail. (a) The U-Net has three scales below the full resolution. The height of a block shows the size of its grid (T × Y × X × Z, below each group), and the number above a block is its output channels. The decoder input [yt, m] goes down to the latent grid, where it is concatenated with the latent z (after one convolution) and mixed by a 1×1×1 convolution and two TBlocks. The up path uses nearest upsampling and gets three skip connections (blue): the two down-path features and the decoder input itself. The last convolution starts at zero, so the first output is zero. (b) The time t becomes one embedding vector e. (c) A TBlock is a residual block of one space convolution and one time convolution; e is added between them. A clip has four axes, so each 4-D convolution is split into a 3-D convolution over space and a 1-D convolution over time.

Source: DiffDecoder and TBlock in research/bubae/diffdec.py; shapes for a clip of 32 frames and 96 depth voxels.