WildAi
d2390eb476
Garbled, parameter-insensitive speech came from a randomized EOS head: 5.3 re-runs _initialize_weights over acoustic_connector and tts_eos_classifier because the vendored _init_weights override had no _is_hf_initialized guard. Guard it; only checkpoint-absent weights are initialized now. - MockCacheLayer exposes both the 4.x and 5.x cache APIs, so the prefilled voice prompt is visible to 5.3's mask builder. - _ensure_cache_has_layers covers the container: offload/prefetch, batch ops, crop, and a copyable lazy prefetch stream. - max_new_tokens is a combined text+speech budget, not latents. - cfg_scale floor of 1.5 for the realtime family. - voice presets resolve against every registered TTS root. - sage excluded from the realtime path: it ignores the attention mask. - bound transformers to >=5.3.0,<5.4, the measured line.
81 lines
3.2 KiB
Python
81 lines
3.2 KiB
Python
"""ComfyUI folder-registration helpers for VibeVoice model assets.
|
|
|
|
The helpers accept ``folder_paths`` explicitly so registration behavior can be
|
|
unit tested without importing the package startup path or starting ComfyUI.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import os
|
|
from typing import Any
|
|
|
|
|
|
TTS_FOLDER_KEY = "tts"
|
|
VOICE_PRESET_FOLDER_KEY = "vibevoice_voices"
|
|
VOICE_PRESET_SUBDIR = os.path.join("VibeVoice", "voices")
|
|
|
|
|
|
def _append_unique_path(paths: list[str], candidate: str) -> bool:
|
|
"""Append a normalized path once, preserving registration order."""
|
|
normalized = os.path.abspath(os.path.normpath(candidate))
|
|
existing = {os.path.abspath(os.path.normpath(path)) for path in paths}
|
|
if normalized in existing:
|
|
return False
|
|
paths.append(normalized)
|
|
return True
|
|
|
|
|
|
def register_vibevoice_folders(folder_paths_module: Any) -> list[str]:
|
|
"""Register the primary TTS root and return all registered TTS roots.
|
|
|
|
Existing ComfyUI registrations are preserved. Calling the helper more than
|
|
once does not add duplicate paths.
|
|
"""
|
|
models_dir = os.fspath(folder_paths_module.models_dir)
|
|
primary_tts_root = os.path.join(models_dir, "tts")
|
|
folder_names_and_paths = folder_paths_module.folder_names_and_paths
|
|
|
|
if TTS_FOLDER_KEY not in folder_names_and_paths:
|
|
supported_exts = set(folder_paths_module.supported_pt_extensions)
|
|
supported_exts.update({".safetensors", ".json"})
|
|
folder_names_and_paths[TTS_FOLDER_KEY] = ([], supported_exts)
|
|
|
|
registered_paths = folder_names_and_paths[TTS_FOLDER_KEY][0]
|
|
_append_unique_path(registered_paths, primary_tts_root)
|
|
return list(registered_paths)
|
|
|
|
|
|
def register_voice_preset_folder(
|
|
folder_paths_module: Any,
|
|
primary_tts_root: str,
|
|
additional_tts_roots: list[str] | None = None,
|
|
) -> list[str]:
|
|
"""Register ``<TTS root>/VibeVoice/voices`` for cached voice prompts.
|
|
|
|
Every TTS root ComfyUI knows about gets a candidate ``voices`` directory,
|
|
not just the primary one. A machine can resolve models across several roots
|
|
at once (the primary ``models/tts`` plus whatever ``extra_model_paths.yaml``
|
|
contributes), and the prompts may sit under any of them. Registering only
|
|
the first root made ``vibevoice_voices`` resolve to a directory that does
|
|
not exist, so the node reported every preset as missing even though the
|
|
prompts were installed under a different root.
|
|
|
|
Directories that do not exist are still registered: ComfyUI populates the
|
|
folder list at startup, and a root may be populated later. Discovery skips
|
|
non-existent paths, so this costs nothing.
|
|
"""
|
|
folder_names_and_paths = folder_paths_module.folder_names_and_paths
|
|
if VOICE_PRESET_FOLDER_KEY not in folder_names_and_paths:
|
|
folder_names_and_paths[VOICE_PRESET_FOLDER_KEY] = ([], {".pt"})
|
|
|
|
registered_paths = folder_names_and_paths[VOICE_PRESET_FOLDER_KEY][0]
|
|
roots = [primary_tts_root, *(additional_tts_roots or [])]
|
|
for root in roots:
|
|
if not root:
|
|
continue
|
|
voice_root = os.path.abspath(
|
|
os.path.join(os.fspath(root), VOICE_PRESET_SUBDIR)
|
|
)
|
|
_append_unique_path(registered_paths, voice_root)
|
|
return list(registered_paths)
|