levzalt
/

Safetensors
File size: 2,478 Bytes
9781faf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
---
license: mit
---
# Audio DiT

Tag-conditioned music generation focused on instrumental game and anime songs. Choose 1-8 tags (recommend to start with 1). Generates 95s songs.
Set a seed to reproduce a generation, or leave it at `-1` for a random seed.

## Run it

Needs [uv](https://docs.astral.sh/uv/). CUDA, MPS or CPU.

```bash
hf download Localsong/Localsong --local-dir Localsong
uv run webui.py  # http://127.0.0.1:7860
```

## Architecture

```
tags ──► AudioDiT ──► 256×1024 latents ──► SAME-L decoder ──► 95 s of stereo audio
         0.84B params            
```

**Latents.** [SAME-L](https://huggingface.co/stabilityai/SAME-L) autoencoder by Stability AI.

**Denoiser.** `audio_dit.py` is a 1-D latent transformer DiT. 
RMSNorm and QK-RMSNorm, RoPE, SwiGLU
Timestep conditioning is adaLN-single (PixArt)
offset. Tags condition via cross-attention.

**Sparse-dense fusion** ([SPRINT](https://arxiv.org/abs/2510.21986)). The blocks
are split into a dense 2-block encoder, a 20-block middle stack, and a dense
2-block decoder. The encoder output is concatenated with the middle output and
projected back down before the decoder:

```
frames ──► encoder ──┬─────────────────────────► f ──┐
                     └─► middle (20 blocks) ──► g ──► linear([f;g]) ──► decoder ──► velocity
```

Training ran the middle stack on a random 25 % of frames and sometimes dropped
it entirely, which makes the shallow encoder→fusion→decoder route a usable weak
model on its own. Inference runs the middle stack on every frame
(`AudioDiT.forward`) and keeps the shallow route as the guidance branch
(`AudioDiT.shallow`).

**Sampling.** Rectified flow: v-prediction, 50 steps Euler shift=2
Guidance alternates between normal CFG and the shallow-path PDG.

## License

MIT, see [LICENSE](LICENSE).

`same_l_decoder.py` is adapted from Stability AI's
[stable-audio-3](https://github.com/Stability-AI/stable-audio-3) (MIT). The
SAME-L weights it loads are released by Stability AI under the Stable Audio
Community License; commercial use of them is subject to
<https://stability.ai/license>.

The model also builds on SPRINT ([2510.21986](https://arxiv.org/abs/2510.21986)),
PixArt-α ([2310.00426](https://arxiv.org/abs/2310.00426)),
DiT ([2212.09748](https://arxiv.org/abs/2212.09748)) and
rectified flow ([2209.03003](https://arxiv.org/abs/2209.03003)).