t8star/Drama-3-Comfy

DRAMA-3 · T8 native models

English | 简体中文

Converted models for ComfyUI-Drama3-T8 and its local voice studio. Built with Fish Audio. Use T8 code 0.2.0 or later for the W4A8 model, compact control head and native single-file aligner. These files use the T8/ComfyUI native layout, not the Transformers loading API.

File	Size	ComfyUI destination
drama3-t8-w4a8.safetensors — W4A8 experts, BF16 dense layers	24.60 GB	models/diffusion_models/
drama3-t8-fp8.safetensors — alternative main model	38.58 GB	models/diffusion_models/
drama3-codec.safetensors	1.57 GB	models/vae/
drama3-line-head.safetensors — compact scene/control head	8.48 KB	models/diffusion_models/
drama3-text-head.safetensors — legacy full-head alternative	1.03 GB	models/diffusion_models/
qwen3_forced_aligner_0.6b_bf16.safetensors — optional word timing	1.84 GB	models/asr/
qwen3-asr-06b.safetensors / qwen3_asr_1.7b_bf16.safetensors — optional ASR	1.56 / 4.08 GB	models/asr/
Confucius4-R2T2-GGUF/ — optional model + matching projector	2.18 GB	models/asr/
FireRedVAD-ONNX/ — optional long-clip segmentation	2.3 MB	models/asr/

Choose one main model plus Codec. Scenes and director revisions also need a compact or full control head. The existing qwen-aligner/ directory remains available as an alignment alternative. Download complete auxiliary folders; R2T2 requires its matching projector and the separately installed native runtime.

All model paths can be changed in nodes or Studio. Studio uses its own models/ directory. Voice cloning conditions new text on reference audio and its matching transcript; no voice training or external inference service is required.

W4A8 reduces expert storage, while memory and speed depend on context, cache and hardware. Core testing uses Windows x64, RTX 5090 Laptop 24 GB VRAM / 64 GB RAM, Torch 2.8 + CUDA 12.8 and Transformers 4.57.1. This is a limited configuration, not a universal accuracy or speed benchmark. W6A8/INT8 are not included in this upload; the code repository provides local conversion tools and their current qualification limits.

DRAMA weights retain the Fish Audio Community License v1.0, derived from fishaudio/drama-3 at e0a9218aa48888270f06ef1ed9ea67622c0e17d2. Qwen ASR/aligner retain Apache-2.0. Confucius and FireRedVAD retain their existing licenses. The prior Qwen3-ASR 1.7B license and notice are preserved. See file checksums, fixed upstream sources and conversion provenance. Existing files are preserved; see the code repository for installation, examples and validation scope.

T8star

Bilibili · YouTube · API · Gallery · Online apps · ComfyUI bundle · Hugging Face