Does each modality need its own diffusion time?

The ablation, at the matched checkpoint

Both runs saved a checkpoint at optimiser step 12,000. This is the only comparison in which nothing but the flag differs; everything below it is context. Edit size = ‖generated − input‖ / ‖target − input‖ on the waveform — 1.00 means the model moved the audio exactly as far as the ground truth did.

Every checkpoint scored on these 84 clips

Pass-through is the floor: return the input untouched. Because the target differs from the input by only one amplified object, a model that does nothing already scores 21.20 dB, so PSNR alone cannot tell you whether an edit happened — read it together with edit size.

• The ablation is a null result so far. At step 12k the shared-t run is +0.43 dB on audio PSNR and −0.52 dB on video PSNR. Both are inside the baseline’s own checkpoint-to-checkpoint spread on this test set (audio PSNR moves 21.07 → 21.68 → 22.00 → 21.67 across 8k/12k/16k/24k; FVD moves 327 → 289 → 297 → 319), so neither sign survives as evidence.
• The one difference that is not noise is that the video dip does not reproduce. Shared-t scores 17.60 dB / SSIM 0.611 at 12k but 18.19 dB / 0.628 at 14.3k — a 0.6 dB swing over 2,300 steps, where the baseline moves by 0.07 dB over 16,000. Read the 12k video row as one noisy point, not as a cost of sharing t.
• The diagnostic the flag was supposed to move did not move. The loss-vs-t probe at the matched step is the same curve in both runs (audio, relative to a predict-nothing baseline: 0.12 / 0.11 / 0.12 / 0.14 at t = 0.02 / 0.05 / 0.1 / 0.2 for independent t, and 0.11 / 0.10 / 0.11 / 0.13 for shared; video is flat near 1.0 in both). Whatever the second timestep embedder buys, it is not a different objective landscape — and it costs 3.67 M parameters, 0.31 % of the trainable model.
• Both runs are far above the earlier σ = 0.01 model (FVD 666, audio 19.81 dB), so the change that mattered was the bridge noise, not how t is drawn.
• Caveat: the ablation run is unfinished. It is at 14.3k of a planned 24k steps, and the baseline is not monotone in training — its best audio point is 16k, not its last. The comparison has to be repeated at 24k before the ablation cell can be reported as settled.