Spatial audio as a visible control signal

Place the sound.

Choose a trained reverb size and dry/wet category, switch the outdoor space between day and night, or route through experimental Dual Delay SFX. The holographic volume is a legible proxy for the caption the model actually learned.

Source

Listener

Space · Moderate

1.90 s decay · source setting ≈1.92 s

Drag to orbit · scroll to zoom

Scene propertiesdrag · reset

Reverb sizeMedium SmallClubMediumCathedral Dry / wetModerate LowModerateHeavy

Estimated RT60 1.90 s

Pre-delay 11 ms

LoRA path On · 1.0

Acoustic scene

Medium room

trained

Controlmode

Reverb sizetrained presets

Dry / wettrained category

Three trained prompt categories; the fader chooses learned wording, not a DSP value.

Generated conditioningcaption + scene data

Listening proof

Six dry / AKUSPACE comparisons

0 / 6 ready Loading examples

Example 01 0:00 / 0:00

Dry AKUSPACE

Both files play in sync — drag the fader to A/B the same moment.

Compare your own audio

Dry reference Processed

Selecting a comparison also restores its trained scene above.

The approach

Spatially aware generation.

A generated video can place a performer in a cathedral, an empty club or an open field while the voice still sounds like a close, dry studio recording. The picture describes a space, but the sound ignores it. That mismatch is one reason otherwise convincing generated footage can still feel synthetic.

AKUSPACE closes that gap by re-generating the audio take with the acoustic character of the scene. The performance returns inside the requested room or place, with its timing preserved for synced video. It does not change the pixels; it makes the picture and sound feel as though they belong to the same environment.

266: paired clips228 train · 38 validation
364: distinct audio files98 dry · 266 processed
14: sourcesacross 7 source types
19: treatment cellsevery cell carries all 14 sources
5: training runsv1 to v5 · v5 ships

Dataset and training Owned recordings · paired transformations · trained prompt levels

Where the material comes from

All training material is owned and was recorded or produced over several years: speaking voices, beats and electronic music, percussion, and acoustic instruments. Treatments use digital reverbs, custom presets, original Eurorack patches and original field recordings. Nothing was scraped or third-party licensed for training.

How it is built

Every item contains the same performance twice: dry and through a real treatment chain. This teaches a transformation instead of an association. Rooms and sound effects use gentle, moderate and heavy; outdoor places use gentle and heavy. The level word stays in the same caption position, and the ComfyUI nodes emit those exact trained strings.

LTX-2.5 Audio LoRA v0.5 · trigger AKUSPACE · checkpoint 11500 · rank 32 · trained with the official LTX Trainer on LTX-2.5. Moderate is the showcase level; heavy ships as experimental.