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.

Needs uv. CUDA, MPS or CPU.

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

Latents. 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). 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.

MIT, see LICENSE.

same_l_decoder.py is adapted from Stability AI's 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), PixArt-α (2310.00426), DiT (2212.09748) and rectified flow (2209.03003).