Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Add VAE tiled decode node for audio. (#12299)

comfyanonymous35183543e004
diff --git a/comfy/sd.py b/comfy/sd.py
index bc63d6ced..bc9407405 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -976,7 +976,7 @@ class VAE:
if overlap is not None:
args["overlap"] = overlap
- if dims == 1:
+ if dims == 1 or self.extra_1d_channel is not None:
args.pop("tile_y")
output = self.decode_tiled_1d(samples, **args)
elif dims == 2:
diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py
index bef723dce..b63dd8e97 100644
--- a/comfy_extras/nodes_audio.py
+++ b/comfy_extras/nodes_audio.py
@@ -94,6 +94,19 @@ class VAEEncodeAudio(IO.ComfyNode):
encode = execute # TODO: remove
+def vae_decode_audio(vae, samples, tile=None, overlap=None):
+ if tile is not None:
+ audio = vae.decode_tiled(samples["samples"], tile_y=tile, overlap=overlap).movedim(-1, 1)
+ else:
+ audio = vae.decode(samples["samples"]).movedim(-1, 1)
+
+ std = torch.std(audio, dim=[1, 2], keepdim=True) * 5.0
+ std[std < 1.0] = 1.0
+ audio /= std
+ vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+ return {"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]}
+
+
class VAEDecodeAudio(IO.ComfyNode):
@classmethod
def define_schema(cls):
@@ -111,16 +124,33 @@ class VAEDecodeAudio(IO.ComfyNode):
@classmethod
def execute(cls, vae, samples) -> IO.NodeOutput:
- audio = vae.decode(samples["samples"]).movedim(-1, 1)
- std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0
- std[std < 1.0] = 1.0
- audio /= std
- vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
- return IO.NodeOutput({"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]})
+ return IO.NodeOutput(vae_decode_audio(vae, samples))
decode = execute # TODO: remove
+class VAEDecodeAudioTiled(IO.ComfyNode):
+ @classmethod
+ def define_schema(cls):
+ return IO.Schema(
+ node_id="VAEDecodeAudioTiled",
+ search_aliases=["latent to audio"],
+ display_name="VAE Decode Audio (Tiled)",
+ category="latent/audio",
+ inputs=[
+ IO.Latent.Input("samples"),
+ IO.Vae.Input("vae"),
+ IO.Int.Input("tile_size", default=512, min=32, max=8192, step=8),
+ IO.Int.Input("overlap", default=64, min=0, max=1024, step=8),
+ ],
+ outputs=[IO.Audio.Output()],
+ )
+
+ @classmethod
+ def execute(cls, vae, samples, tile_size, overlap) -> IO.NodeOutput:
+ return IO.NodeOutput(vae_decode_audio(vae, samples, tile_size, overlap))
+
+
class SaveAudio(IO.ComfyNode):
@classmethod
def define_schema(cls):
@@ -675,6 +705,7 @@ class AudioExtension(ComfyExtension):
EmptyLatentAudio,
VAEEncodeAudio,
VAEDecodeAudio,
+ VAEDecodeAudioTiled,
SaveAudio,
SaveAudioMP3,
SaveAudioOpus,