0.1.71

Higgs_v3-TTS-ComfyUI v0.1.71

Fixed

• Fixed AIMDO unload cleanup throwing ignored KeyError: device(type='cuda', index=0) after successful generation.
• Higgs unload now matches the Moss TTS unload behavior by preserving dynamic_pins for ComfyUI/AIMDO cleanup.
• The loader still clears dynamic_vbars, moves weights to meta, runs garbage collection, and empties accelerator cache on hard unload.

Changed

• Bumped package version to v0.1.71.
• Updated README and README_zh version badges to v0.1.71.

Notes

• This is a cleanup/hotfix release for AIMDO unload behavior.
• No node interface changes.
• No workflow changes required.

0.1.7

Higgs_v3-TTS-ComfyUI v0.1.7

Added

• Added real ComfyUI/AIMDO DynamicVRAM support for the native Higgs v3 model and audio codec.
• Higgs model and codec weights now load CPU-first, then register through ComfyUI model management.
• Added Comfy-castable wrappers for Higgs/Qwen/codec modules so AIMDO can stage and page weights instead of force-preloading them.
• Added AIMDO-aware Whisper ASR staging:
  • Whisper loads on CPU first when AIMDO DynamicVRAM is active.
  • Whisper modules are converted for Comfy dynamic casting before registration.
  • Whisper now participates in ComfyUI memory management more correctly.

Fixed

• Fixed HiggsAudioV2TokenizerModel crashing during AIMDO registration because its .device property was read-only.
• Fixed ugly WhisperForConditionalGenerationComfyWritableDevice / codec class names in logs while keeping the writable device shim.
• Fixed partial-load cleanup so if codec registration fails after the main model loads, the already-loaded model is detached and cleaned up instead of lingering.
• Improved hard unload integration with ComfyUI native unload calls.
• Updated Whisper registration warning text to refer to ComfyUI memory management.

Changed

• Replaced the older custom/fake Higgs memory tracking path with ComfyUI’s real model patcher flow.
• Updated README and README_zh to document AIMDO DynamicVRAM behavior.
• Updated package metadata to v0.1.7.

Notes

• AIMDO DynamicVRAM can reduce live VRAM pressure by paging weights, but the recommended target remains 11 GB+ VRAM until lower-VRAM workflows are tested thoroughly.

0.1.5

Higgs_v3-TTS-ComfyUI v0.1.5

Fixed

• Emotion tags no longer automatically carry into later longform chunks.
• Explicit emotion tags inside later chunks continue to work normally.
• Reduced cloned-speaker identity drift caused by strong emotions being inserted at the beginning of every chunk.
• Positive fixed seeds are now reused unchanged across all longform chunks.
• Reference audio and reference text remain consistent across Voice Clone chunks.

Improved

• Added continuous ComfyUI node progress for Generate and Voice Clone.
• Added continuous progress across longform chunks.
• Added continuous progress across Multi-Speaker turns.
• Progress no longer resets between chunks or speaker turns.
• Natural early completion correctly advances the node progress bar to 100%.
• Style, speed, pitch, and expressiveness tags continue carrying into later chunks.
• Updated English and Chinese documentation.

0.1.4

Higgs_v3-TTS-ComfyUI v0.1.4

Added

• Added a live tqdm terminal progress bar during audio-token generation.
• Progress now displays percentage, elapsed time, ETA, and token generation rate.
• Natural early completion adjusts the total to the actual generated token count and finishes at 100%.
• Added tqdm as an automatically installed lightweight dependency.
• Updated English and Chinese documentation.

Example

Higgs v3 audio tokens: 64%|████████████████████████▋| 1310/2048 [00:39<00:21, 34.52tok/s]

0.1.3

v0.1.3

Changed

• Removed the emotion, style, speed, pitch, and expressiveness dropdowns from Generate, Voice Clone, and Multi-Speaker nodes.
• Inline control tags are now the intended control path for emotion, style, prosody, pauses, and sound effects.
• Updated README and Chinese README to document inline tags instead of removed dropdown controls.
• Updated example workflow JSON and preview image for the new node layout.
• Bumped package version to v0.1.3.

Fixed

• Hard unload now skips moving Higgs weights to CPU RAM first and sends tensors directly to meta where supported.
• This reduces the “VRAM unload becomes sticky system RAM” behavior during hard unload/model replacement.
• Longform chunking still preserves inline delivery tags across chunks.

0.1.2

v0.1.2

Fixed

• Removed the forced minimum audio-token guard that suppressed Higgs' natural EOC/stop token.
• Single-pass generation now stops when the model emits its real stop token instead of being forced to continue.
• This should reduce repeated speech, unstable tails, crackles, and codec artifacts caused by over-generation.

0.1.1

v0.1.1

Added

• Added native Higgs v3 TTS ComfyUI nodes:
  • Higgs v3 Load Model
  • Higgs v3 Generate
  • Higgs v3 Voice Clone
  • Higgs v3 Multi-Speaker
  • Higgs v3 Whisper Transcribe
• Added dynamic multi-speaker inputs for 2-6 speakers on newer ComfyUI builds.
• Added longform chunking with sentence/pause-aware splitting.
• Added inline support docs for emotion, style, prosody, and sound effect tags.
• Added ComfyUI/AIMDO memory tracking for Higgs model and codec tensors.
• Added pyproject.toml metadata for release/registry use.

Changed

• Removed fp16 from Higgs dtype options because it can produce non-finite audio.
• Kept dtype dropdown with auto and bf16.
• Changed auto dtype to use bf16 on supported CUDA and fp32 otherwise.
• Removed fp16 from Whisper dtype options too; Whisper now exposes auto, bf16, and fp32.
• Changed default words_per_chunk from 100 to 45 for safer generation with max_new_tokens=2048.
• Generate longform now uses chunk 1 as an internal voice reference for later chunks when no external reference audio is connected.
• Voice Clone longform now reuses the original reference_audio and reference_text for every chunk.
• When longform chunking is off, Generate/Clone now run a direct single pass instead of going through the chunking path.
• Improved hard unload behavior by unregistering Comfy model patchers, clearing AIMDO state, moving weights to meta, breaking bundle references, running GC, and emptying accelerator caches.
• Updated README and README_zh with Transformers compatibility notes:
  • Supported: Transformers 5.3.0 through 5.5.0
  • Recommended: Transformers 5.5.0
• Documented the native workaround for the Transformers version issue: direct Qwen3/Higgs module construction, safetensors weight mapping, and tokenizer config normalization.
• Documented expected VRAM use: about 11 GB with bf16 on CUDA.
• Updated README SFX/prosody formatting so <|...|> tags render correctly on GitHub.

Fixed

• Fixed broken README Markdown tables where <|sfx:...|> and <|prosody:...|> tokens were split by pipe characters.
• Fixed misleading chunk 1/1 behavior/logging when chunking was disabled.
• Fixed multi-speaker input ordering so speaker audio/reference text fields are paired correctly.
• Fixed dynamic multi-speaker options to support 2, 3, 4, 5, and 6 speakers.
• Fixed loader notice wording from “visualization” to “memory tracking.”
• Added an early stop guard to reduce cases where generation ends far too early for the amount of text.

Notes

• SageAttention remains available, but may be slower than SDPA/FlashAttention for Higgs v3’s token-by-token TTS generation path.
• max_new_tokens=2048 is roughly 25-30 seconds of audio. Use longform chunking or raise max_new_tokens for longer text.
• ComfyUI offload can still keep weights in system RAM by design. Hard unload happens when changing model/dtype/device/attention settings.

0.1.0

Initial Release