Note that DiffRhythm-base requires a minimum of 8G of VRAM

[nitinmukesh]

Thank you for releasing the code and models. Very good quality.

Please clarify 8G VRAM for what duration of song.

I tested with 1.35m base model and it need 14.5 GB+

[NZqian]

DiffRhythm/infer/infer.py

Line 52 in ba46074

output = decode_audio(latent, vae_model, chunked=False)

Enable chunk infer of VAE can greatly reduce VRAM cost. Thank you for noticing, we will clarify in the readme.

[nitinmukesh]

@NZqian

You rock. Now it's only 6.6 GB and inference took only few sec. Earlier it was more than 5 mins.