Add MOSS community full-checkpoint support
Implementation details:
- Add the LAION MOSS-TTS v1.5 Voice Acting 8B community model to the engine dropdown and downloader
- Discover compatible local MOSS full checkpoints and resolve architecture from config.json
- Enable the LAION Delay checkpoint in the existing LoRA training pipeline
- Reject unsupported local MOSS layouts instead of assuming Local Transformer
- Address issue #340
Release IndexTTS 2.5 model-version support
Implementation details:
- Add the official multilingual 2.5 backend and pinned model snapshot
- Add explicit language, duration-factor, and text-normalization controls
- Preserve IndexTTS 2.0 compatibility and emotion-control features
- Fix Text and SRT processor invalidation when switching model versions
- Separate generated-audio cache entries by model and 2.5 parameters
Release DramaBox LoRA training support
Implementation details:
- Add DramaBox dataset preparation, training configuration, and integrated trainer nodes
- Add official audio-branch preprocessing and IC-LoRA training support
- Add DramaBox LoRA adapter loading with runtime reuse and cache-aware strength changes
- Generalize shared training clip staging and dataset-row handling
- Add DramaBox training documentation and example workflow
Fix MOSS Dataset Prep workflow compatibility
Technical details:
- Restore the original serialized order of all existing optional widgets
- Append recursive_folder_scan after legacy widget values
- Prevent saved workflows from shifting validation and codec settings
- Address issue #340
Handle incompatible FlashAttention installs in F5-TTS
Technical details:
- Validate optional FlashAttention submodules before enabling the backend
- Keep F5-TTS on PyTorch attention when FlashAttention is partial or incompatible
- Improve the explicit FlashAttention backend error
- Address issue #334
Allow installer to continue without optional audio libraries
Technical details:
- Treat PortAudio and libsamplerate checks as non-fatal warnings on Linux and macOS
- Allow Fish Audio S2 runtime installation to run in headless environments
- Correct Fedora dependency guidance
- Address issues #330 and #332
Release Character Alias Manager improvements
Technical details:
- Add per-user grouped alias persistence and management API
- Add character preview and missing-transcript metadata
- Integrate alias management with Character Voices and Multiline TTS Tag Editor
- Improve drag feedback, waveform rendering, zoom behavior, discovery refresh, and logging
Release unified voice design and MOSS sound effects
Implementation details:
- Add MOSS-TTS v1.5 model selection and LoRA training support
- Add MOSS-SoundEffect v1 through the MOSS engine
- Add dedicated MOSS-SoundEffect v2 diffusion engine
- Standardize Voice Designer and Sound Effects provider routing
- Add reusable Save Character Voice persistence
- Refactor Character Voices discovery and compact preview layout
- Remove the legacy Qwen-specific voice designer node
- Add example workflows, guides, model metadata, and cleaner download logging
Release Fish and VibeVoice compatibility repairs
Technical details:
- Restore the removed TorchAudio backend probe for Fish Audio S2
- Resolve local VibeVoice paths in the parent ComfyUI process
- Pass absolute model paths to isolated VibeVoice workers
- Add focused runtime regression tests
Fix Fish Audio S2 and Dots installer compatibility
Technical details:
- Restore Fish source subpackages omitted by the upstream wheel metadata
- Allow the tested Dots source install on Python 3.13 without changing dependencies
Refine Character Voices waveform and discovery
Implementation details:
- Add normalized cached waveform rendering to the trim timeline
- Add bounded playback progress with a smooth playhead
- Anchor the trim transcription warning outside layout flow
- Stop node schema queries from forcing voice discovery rescans
- Deduplicate background discovery refreshes across instances
Follow-up to #318
Address CosyVoice3 ROCm precision handling (issue #323)
Technical details:
- Run the bundled Qwen LLM with BF16 autocast on ROCm
- Keep the CosyVoice3 flow model and HiFT vocoder in FP32
- Disable unsupported TensorRT loading on ROCm
- Forward LLM worker failures instead of decoding empty tokens
Fix IndexTTS-2 emotion vector import interface
Implementation details:
- Reuse the shared emotion configuration modal for import and export
- Allow JSON to be pasted and edited before an explicit Apply action
- Validate every imported value before changing any emotion control
- Remove clipboard-first automatic import behavior
Fix Fish Audio S2 installer runtime validation
- Validate fish_speech.inference_engine instead of distribution metadata alone
- Repair incompatible Fish installations automatically
- Pin the tested Fish S2 v2.0.0-beta source revision
- Verify the runtime import after installation
Release Multiline TTS Tag Editor inline emotion controls
Implementation details:
- Add engine-aware one-click and hold-drag tag swapping
- Support combined IndexTTS audio references and vector or text emotion controls
- Improve quoted emotion routing and long inline tag wrapping
- Fix tag parsing, character switching, and newline insertion behavior
Address issue #312 with MPS-safe RVC tensor typing
Technical details:
- Keep RVC pitch indices integer on Apple MPS devices.
- Normalize phone features to the linear layer weight dtype.
- Normalize pitch values to torch.long at the encoder boundary.
- Preserve CUDA half-precision compatibility.
Address issue #311 with resilient MelBand resampling
Technical details:
- Route MelBand input conversion through the shared safe resampler.
- Fall back to torchaudio when librosa/Numba resampling is incompatible.
- Correct the PROJECT_INDEX path for bump instructions.
- Make version tooling use UTF-8 on Windows.
Add Fish Audio S2 Pro engine and integration workflow
Implementation details:
- Integrate the official Fish S2 runtime with isolated loading and model selection
- Add native multi-speaker and custom character-switching generation paths
- Add inline instruction translation, language prompting, caching, compilation, and optional quantization
- Add unified Text and SRT workflow coverage with documentation and a workflow cover
Add Granite ASR 4.1 diarization and timestamp refinements
Implementation details:
- normalize Granite ASR plus diarization output to suite-native [Speaker] tags
- fix Granite ASR plus native timestamp truncation by auto-raising low generation token budgets
- move Unified ASR Transcribe diarization input to preserve old workflow compatibility
- update Granite ASR YAML and README documentation and regenerate engine tables
- improve cross-platform test runner path handling and Granite regression coverage
Fix isolated runtime bootstrap fallback for Unified ASR on portable ComfyUI
Technical details:
- Fall back from python -m venv to python -m virtualenv when embedded Windows builds cannot create a venv
- Preserve base site-packages inheritance so isolated runtimes still reuse ComfyUI heavy dependencies
- Surface primary and fallback bootstrap stderr/stdout for easier runtime setup debugging
- Address issue #304
Add Dots TTS engine and refine language handling
Implementation details:
- integrate official Dots TTS text and SRT workflows
- improve Dots runtime lifecycle, caching, and progress reporting
- fix Dots language override behavior and refresh generated metadata
Prepare v5.0.0 release
Major changes:
- Resume Higgs Audio v3 on the main Transformers 5 environment
- Add isolated runtime support for fragile legacy engines
- Improve Multiline TTS Tag Editor inline validation, conversion, and persistence
- Refresh README, workflows, screenshots, and engine documentation
Change RVC custom chunking to opt-in
Technical details:
- Add an RVC engine toggle for extra outer chunking on top of native segmentation
- Default RVC to native segmentation instead of the unified outer chunk layer
- Report active RVC chunking mode in Voice Changer conversion info
Add initial MOSS-TTS LoRA training support
Implementation details:
- add unified MOSS Delay 8B LoRA training flow and dataset preparation nodes
- add MOSS training docs and example workflow for clip-based dataset building
- add 4-bit MOSS training option for lower-VRAM LoRA training
- allow MOSS direct TTS inference without forced narrator fallback
Fix Text to SRT Builder Chinese punctuation spacing
Technical details:
- Add CJK-aware SRT text joining for subtitle rendering
- Tokenize CJK plain-text timing estimates by character instead of whitespace
- Preserve Chinese punctuation and quote spacing without standalone punctuation leakage
- Add regression tests for Chinese punctuation, quote spacing, and English spacing
Add MOSS LoRA loading support and compatibility validation
Implementation details:
- Add managed MOSS LoRA discovery and Hugging Face install flow under models/TTS/moss_tts/loras
- Wire PEFT adapter loading through the MOSS engine, adapter, cache, generation info, and unified model interface
- Validate adapter base-model compatibility before loading to prevent huge PEFT mismatch errors
- Raise real ComfyUI popup errors for MOSS LoRA/base model mismatches instead of generic engine creation failures
- Patch bundled MOSS Delay and TTSD classes for PEFT generation compatibility
- Upgrade installer requirement handling so outdated packages like peft are upgraded automatically
- Document the MOSS LoRA folder contract and update MOSS metadata
Fix MOSS model selector stability across local/friendly variants
Technical details:
- Keep local:* model labels visible when local folders exist
- Accept both friendly and local:* values for backward compatibility
- Ensure native mode resolves to TTSD model variant for execution
Fix Qwen3-TTS auto attention fallback for broken flash-attn installs
Technical details:
- Validate flash_attn package metadata before selecting flash_attention_2 in auto mode
- Fall back to sdpa when flash-attn is missing or partially installed on Windows
- Add runtime retry to sdpa when auto-selected flash_attention_2 fails during model load
Address issue #294
Fix Echo-TTS character voice file-path fallback and compressed audio loading
Technical details:
- Preserve Character Voices and dropdown audio_path fallback in Echo-TTS voice mapping
- Allow Echo-TTS narrator and character fallback to use file paths when decoded tensors are unavailable
- Expand safe audio loading fallback order to support OGG and other compressed formats when torchaudio fails
diodiogod