Text-to-Speech
Transformers
Safetensors
PyTorch
English
Chinese
breeze
text-generation
cuda
english
chinese
speech-generation
voice-clone
voice-design
voice-direction
comfyui
int8
convrot
quantization
quantised
hadamard
Instructions to use drbaph/Breeze-TTS-2-comfyui with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use drbaph/Breeze-TTS-2-comfyui with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-to-speech", model="drbaph/Breeze-TTS-2-comfyui")# Load model directly from transformers import AutoModelForSeq2SeqLM model = AutoModelForSeq2SeqLM.from_pretrained("drbaph/Breeze-TTS-2-comfyui", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Breeze TTS 2 β ComfyUI derivatives
Converted from BreezeBlue/Breeze-TTS-2 for the ComfyUI-Breeze-TTS-2 node pack.
Contents
| File | Description |
|---|---|
Breeze-TTS-2-bf16.safetensors |
The official checkpoint merged from 2 shards into a single file (bit-exact, bf16, 6.49 GiB). |
Breeze-TTS-2-int8-convrot.safetensors |
All 462 transformer projection linears INT8 ConvRot (backbone + text encoder + depth decoder), comfy-kitchen per-channel per-row scales. Smallest build. |
Breeze-TTS-2-int8-hybrid.safetensors |
Backbone + text encoder INT8 ConvRot; the depth decoder (the per-frame autoregressive hot loop) stays bf16. Recommended INT8 build: bf16-level speed with β27 % VRAM. |
Breeze-TTS-2-int8-text-encoder.safetensors |
Text encoder only INT8; everything decode-time stays bf16. |
audio_tokenizer/ |
Unmodified Qwen3-TTS 12 Hz audio tokenizer (codec) used for reference-audio encoding and speech decoding. |
quantization_report_*.json |
Per-tensor dequantization error statistics for each INT8 build. |
benchmark_report.json |
RTF / VRAM / log-mel / Whisper-WER comparison across all four builds. |
Benchmarks (RTX 5090, ComfyUI loader, sdpa, 24 kHz output)
| Build | avg RTF (lower = faster) | Peak VRAM | Whisper WER vs prompt* |
|---|---|---|---|
| bf16 | 5.68 | 7.51 GiB | baseline |
| int8-convrot | 9.08 (+60 %) | 5.21 GiB (β31 %) | matches bf16 |
| int8-hybrid | 5.64 (parity) | 5.53 GiB (β27 %) | matches bf16 |
| int8-text-encoder | 5.34 | 6.82 GiB (β9 %) | matches bf16 |
*All builds including bf16 share the same few ASR-level blurs on the clone prompt; design/direction prompts transcribe identically to bf16 on the hybrid builds.
The depth decoder runs 15 skinny single-token GEMMs per 12.5 Hz frame (β1260 linear calls/frame), where int8 per-call overhead (quantize + dispatch, ~0.1 ms/call on Windows) outweighs the GEMM savings β that is why full INT8 regresses and the hybrid does not.
License
- Weights: the Breeze TTS 2 model weights, these derivative files, and self-hosted outputs are
governed by the BreezeBlue Research and Non-Commercial License (see
license_link). Commercial use requires written authorization from RESONIA, INC. - Node-pack code: ComfyUI-Breeze-TTS-2 is Apache-2.0. The Breeze TTS source code is also Apache-2.0.
- Downloads last month
- 27
Model tree for drbaph/Breeze-TTS-2-comfyui
Base model
BreezeBlue/Breeze-TTS-2