README.md

---
license: other
license_name: ltx-2-community-license-agreement
license_link: https://github.com/Lightricks/LTX-2/blob/main/LICENSE.md
base_model: Lightricks/LTX-2.5
pipeline_tag: audio-to-audio
tags:
  - ltx
  - ltx-2.5
  - lora
  - audio-to-audio
  - audio
  - spatial-audio
  - room-acoustics
  - reverb
  - ambience
  - comfyui
---

# AKUSPACE — LTX-2.5 Audio LoRA v0.5

AKUSPACE is a spatially aware Audio LoRA for LTX-2.5. It gives generated or reference audio the prompted acoustic character of a room, outdoor environment or experimental sound treatment while preserving—or generating—audio in sync with the picture.
AKUSPACE supports two synchronized-video paths. In the audio-to-audio finishing route, a video’s generated soundtrack is re-rendered through AKUSPACE and remuxed onto the same frames. In the native joint workflows, LTX generates the picture and reverberant synchronized audio together, using either a one-pass graph or a two-stage HQ graph.

**[Explore the interactive audiovisual study →](https://akuspace.pages.dev)**
· [ComfyUI nodes](https://github.com/koshimazaki/ComfyUI-Koshi-Nodes)

![The AKUSPACE control surface: a scene volume showing source, listener and the
selected acoustic space, with the trained decay time read out
beneath.](assets/akuspace-viewer.png)

Trigger word: `AKUSPACE`

## Controls

| Mode | Options | Levels |
|---|---|---|
| **Space** — rooms a sound sits *in* | small room, medium room, empty club, cathedral | `gentle` · `moderate` · `heavy` |
| **Place** — environments a sound sits *among* | outdoor day, outdoor night | `gentle` · `heavy` |
| **Sound effects** — processing a sound goes *through* | Dual Delay | `gentle` · `heavy` |

Outdoor places have two levels rather than three: an ambience bed is a separate
recording rather than a reverb tail, so it scales down but not up. 

<video controls playsinline preload="metadata" width="100%">
  <source
    src="https://cdn-uploads.huggingface.co/production/uploads/6430969f2a7fbdfb52582807/hHVrCAg86lcYlsbg7nlYz.mp4"
    type="video/mp4"
  >
  Your browser does not support video playback.
</video>

## Use

Supply the dry audio as the **audio-to-audio reference** and use the complete
trained caption. The level word sits between the space and its character:

```text
AKUSPACE female spoken voice through synthetic cathedral reverb, moderate wide diffuse reflections and a long decaying tail, no background ambience
```

```text
AKUSPACE electronic rhythm loop in a small bathroom-like room, gentle bright close reflections and a short 0.67-second reverb decay, no background ambience
```

Partial captions work less well than complete ones — the trailing clauses were
present in every training caption. Room captions carry a decay time where
applicable; cathedral, the granular effect and both outdoor places have no numeric
decay.

Suggested settings: **24 steps, CFG 1-2** adjust to CFG4 for higher volume and detail.

## Supported workflows

| Input | Output |
|---|---|
| **Audio** | Audio-to-audio treatment for an existing recording |
| **Text + audio** | Text-to-video with AKUSPACE-treated synchronized audio |
| **Image + audio** | Image-to-video with AKUSPACE-treated synchronized audio |



> **Turn prompt enhancement OFF.** The rewriter paraphrases away the trigger word
> and the level word, which are exactly the tokens the control surface depends on.

This adapter transforms audio it is given. Text-to-audio with the LoRA loaded and
no reference present produces near-silence — there is nothing to transform.

**Video.** The proven route is dry voice → audio-to-audio space pass →
image+audio-to-video, with the treated audio held fixed while the image conditions
the first frame. The adapter targets the audio branches; video generation remains
handled by the base model.

## Evaluation

Envelope cross-correlation of the six published examples against their own dry
references, measured on this checkpoint at 2 ms resolution, gives a **16–24 ms
offset** — a small consistent lead rather than per-clip drift, inside lip-sync
tolerance.

<details>
<summary>Full timing table, and how to read it</summary>

| Example | Offset | Peak r |
|---|---|---|
| Voice · small room | −20 ms | 0.96 |
| Voice · cathedral | −22 ms | 0.97 |
| Beat · medium room | −16 ms | 0.34 |
| Beat · empty club | −20 ms | 0.30 |
| Beat · outdoor day | −24 ms | 0.32 |
| Beat · dual delay | −22 ms | 0.34 |

These are the six published examples, not an exhaustive sweep of every
configuration.

Read the correlation column with care: it measures how much the amplitude envelope
*changed*, and reverb changes it by design — most of all on transient-dense
material, where the tail fills the gaps between hits. The low figures on percussion
are the effect working, not a defect. It is a timing measure, not a quality
measure; judge quality by ear.

</details>

## Training

Base `Lightricks/LTX-2.5` (22B dev transformer, bf16), trained with the official
LTX Trainer on 266 paired clips. Trained to 12,300 steps; **this release ships step
11,500**, chosen by ear from test renders.

Every item is the same performance twice — once dry, once through a real acoustic
treatment — so the model learns a transformation rather than an association. All
training material is owned and was recorded or produced over several years:
multiple speaking voices, beats and electronic music, percussion, and acoustic
instruments. Treatments come from digital reverbs, custom presets, original
Eurorack modular patches and original field recordings used for the outdoor beds.
Nothing was scraped, and no third-party licensed audio was used.

<details>
<summary>Training configuration and dataset figures</summary>

| | |
|---|---|
| Rank / alpha | 32 / 32, dropout 0 |
| Target modules | `audio_attn1`, `audio_attn2`, `audio_ff` |
| Optimiser | AdamW, 2e-4, batch 1, bf16, seed 42 |
| Paired clips | 266 — 228 train / 38 validation |
| Distinct audio files | 364 — 98 dry / 266 processed |
| Sources | 14, across 7 source types |
| Treatment cells | 19, every cell carrying all 14 sources |
| Training runs | 5 (v1→v5); v5 ships |
| Clip format | 6.000 s, 48 kHz, −3 dBFS ceiling |
| Held out | one beat and one male voice source |

Sources were rendered as a deliberate grid so that no space is defined by one voice
and no voice is defined by one space. The full config is in
`config/a2a_v5_ltx25.yaml`.

</details>

## Limitations

- Requires an audio reference; it will not generate a space from text alone.
- This is generative transformation, not physically accurate acoustic simulation.
- Audio is regenerated rather than filtered, so wording, timing, pitch or timbre
  can shift.
- Outdoor ambience can mask a quiet source.
- The level axis is a learned caption control, not a calibrated wet/dry percentage.

## Files

| Path | What it is |
|---|---|
| `akuspace-ltx25-v0.5.safetensors` | The adapter — step 11500, bf16, rank 32 |
| `config/a2a_v5_ltx25.yaml` | The training config this run used |
| `examples/` | The six A/B pairs, 320k MP3 |

`dry-voice.mp3` pairs with the two voice examples; `dry-beat.mp3` with the four
beat examples. Rooms are at `moderate`, outdoor at `gentle`, dual delay at `heavy`.
No gain or normalisation was applied on export, so the level relationship between
dry and processed is the one the model produced. Hear them crossfaded at
[akuspace.pages.dev](https://akuspace.pages.dev).

The published examples use a synthetic TTS voice and an AI-generated beat as dry
sources — deliberately, because a replacement voice that needs seating in a scene
is the realistic use case. The training grid itself is my own recordings.

## Licence

Released under the **LTX-2.x Community License** terms that govern the base model.
Confirm the current terms at
https://github.com/Lightricks/LTX-2/blob/main/LICENSE.md before commercial use —
the community licence covers entities under $10M annual revenue, with separate
commercial agreements above that threshold.