ComfyUI v0.12.3

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:

comfy/text_encoders/ace15.py

@@ -3,6 +3,7 @@ import comfy.text_encoders.llama
 from comfy import sd1_clip
 import torch
 import math
+import yaml
 import comfy.utils
 
 
@@ -125,14 +126,43 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_06b", tokenizer=Qwen3Tokenizer)
 
+    def _metas_to_cot(self, *, return_yaml: bool = False, **kwargs) -> str:
+        user_metas = {
+            k: kwargs.pop(k)
+            for k in ("bpm", "duration", "keyscale", "timesignature", "language", "caption")
+            if k in kwargs
+        }
+        timesignature = user_metas.get("timesignature")
+        if isinstance(timesignature, str) and timesignature.endswith("/4"):
+            user_metas["timesignature"] = timesignature.rsplit("/", 1)[0]
+        user_metas = {
+            k: v if not isinstance(v, str) or not v.isdigit() else int(v)
+            for k, v in user_metas.items()
+            if v not in {"unspecified", None}
+        }
+        if len(user_metas):
+            meta_yaml = yaml.dump(user_metas, allow_unicode=True, sort_keys=True).strip()
+        else:
+            meta_yaml = ""
+        return f"<think>\n{meta_yaml}\n</think>" if not return_yaml else meta_yaml
+
+    def _metas_to_cap(self, **kwargs) -> str:
+        use_keys = ("bpm", "duration", "keyscale", "timesignature")
+        user_metas = { k: kwargs.pop(k, "N/A") for k in use_keys }
+        duration = user_metas["duration"]
+        if duration == "N/A":
+            user_metas["duration"] = "30 seconds"
+        elif isinstance(duration, (str, int, float)):
+            user_metas["duration"] = f"{math.ceil(float(duration))} seconds"
+        else:
+            raise TypeError("Unexpected type for duration key, must be str, int or float")
+        return "\n".join(f"- {k}: {user_metas[k]}" for k in use_keys)
+
     def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
         out = {}
         lyrics = kwargs.get("lyrics", "")
-        bpm = kwargs.get("bpm", 120)
         duration = kwargs.get("duration", 120)
-        keyscale = kwargs.get("keyscale", "C major")
-        timesignature = kwargs.get("timesignature", 2)
-        language = kwargs.get("language", "en")
+        language = kwargs.get("language")
         seed = kwargs.get("seed", 0)
 
         generate_audio_codes = kwargs.get("generate_audio_codes", True)
@@ -141,16 +171,20 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
         top_p = kwargs.get("top_p", 0.9)
         top_k = kwargs.get("top_k", 0.0)
 
+
         duration = math.ceil(duration)
-        meta_lm = 'bpm: {}\nduration: {}\nkeyscale: {}\ntimesignature: {}'.format(bpm, duration, keyscale, timesignature)
-        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n{}\n<|im_end|>\n<|im_start|>assistant\n<think>\n{}\n</think>\n\n<|im_end|>\n"
+        kwargs["duration"] = duration
 
-        meta_cap = '- bpm: {}\n- timesignature: {}\n- keyscale: {}\n- duration: {}\n'.format(bpm, timesignature, keyscale, duration)
-        out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, meta_lm), disable_weights=True)
-        out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, ""), disable_weights=True)
+        cot_text = self._metas_to_cot(caption = text, **kwargs)
+        meta_cap = self._metas_to_cap(**kwargs)
 
-        out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric{}<|endoftext|><|endoftext|>".format(language, lyrics), return_word_ids, disable_weights=True, **kwargs)
-        out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}# Metas\n{}<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs)
+        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n# Lyric\n{}\n<|im_end|>\n<|im_start|>assistant\n{}\n<|im_end|>\n"
+
+        out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, cot_text), disable_weights=True)
+        out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, "<think>\n</think>"), disable_weights=True)
+
+        out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric\n{}<|endoftext|><|endoftext|>".format(language if language is not None else "", lyrics), return_word_ids, disable_weights=True, **kwargs)
+        out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}\n# Metas\n{}\n<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs)
         out["lm_metadata"] = {"min_tokens": duration * 5,
                               "seed": seed,
                               "generate_audio_codes": generate_audio_codes,

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,

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.12.2"
+__version__ = "0.12.3"

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.12.2"
+version = "0.12.3"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"