LatentFT
FlowEdit ↗Who you are
The same token as the FlowEdit page. Remembered in this browser only.
What this is
LatentFT (Wang & Huang, ICLR 2026) encodes about 6 seconds of audio into a latent: 80 channels that change 86 times a second. Take the Fourier transform of each channel along time and you get its rate axis: how fast things change, from 0 to 43 Hz. Slow rates are bars and phrases, fast rates are 16th notes, rolls and attacks.
The page splits that axis into the paper's 8 bands, each an octave wide except the lowest two. You choose which bands to keep. The decoder is generative: whatever you remove, it fills back in with something plausible, so this is not a filter on the sound. Band 1 also carries each channel's average, which is the overall timbre and level.
More detail
The output comes from a BigVGAN vocoder at 22.05 kHz, so nothing above 11 kHz survives. "What the model hears" is your clip through that vocoder with no model in between: the best any render can sound.
In Blend, each clip's latent is cut to its own bands and the sampler follows both at once, weighted equally. Measured on two songs, the rhythm follows whichever clip supplies bands 5 to 8 (8th notes and faster). With clip A's bands 1–4 and B's bands 5–8, the onset pattern matched B at 0.86 and A at 0.10. The reverse matched A at 0.95 and B at 0.07.
Every render of one analysis starts from the same noise, which is why neighbouring renders crossfade smoothly. Mono uses Mason Wang's paper checkpoint on a downmix. Stereo uses stereo-v1, a fine-tune that sees left and right together.
1 · Model
2 · Clips
3 · Timescale EQ
Click bands to keep or remove them. Bars show how much of the clip's latent energy sits in each band (after Analyze).
3 · Blend
Click a band to take it from the other clip.
Position k takes the k slowest bands from the first clip named on the Swap button and the rest from the other: 0 is all of one clip, 8 all of the other.
4 · Listen
Everything loops the same 5.9-second window, in sync, so switching keeps your place.