ComfyUI-Drama3-T8

ComfyUI-Drama3-T8

English | 简体中文

Native Fish Audio DRAMA-3 nodes and a local voice studio by T8star. Built with Fish Audio.

Models · Releases · Example workflows · Issues

Features

• Voice cloning with multiple references and up to 16 speaker IDs; inline directions and multi-character scenes.
• Director revisions using actual audio history; infill editing; long text, SRT and streamed text with lookahead.
• Codec encoding/decoding, reusable references, session save/load/branch and audio chunk previews.
• Local Qwen3-ASR, Qwen ForcedAligner and Confucius R2T2, with editable model paths.
• 17 native ComfyUI V3 nodes using standard AUDIO. Inference, codec and R2T2 run inside the host process. The separate white-and-pink studio shares the inference core.

Install in ComfyUI

Install ComfyUI-Drama3-T8, publisher t8star, through ComfyUI-Manager, or install manually:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-Drama-3.git
cd Comfyui-Drama-3
python -m pip install -r requirements.txt
python install.py

Use your ComfyUI Python, then restart ComfyUI. Import a workflow from workflows/ and select your models and reference audio. Requires ComfyUI 0.40.0+ and Python 3.10+. install.py downloads and verifies the separately released Windows x64 / CUDA 12.8 R2T2 native runtime; it never downloads models. Set DRAMA_SKIP_R2T2_INSTALL=1 to skip this optional backend, or DRAMA_R2T2_NATIVE_DIR to use an existing runtime. Qwen ASR works independently.

Transformers compatibility

Transformers 4.57.1 is supported; an existing compatible 4.x installation can be kept. The nodes accept transformers>=4.57.1,<6 and tokenizers>=0.22,<0.24.

| Validated Transformers | Tokenizers | Hugging Face Hub | |---|---|---| | 4.57.1 | 0.22.2 | 0.36.2 | | 5.19.0 | 0.23.2 | 1.33.0 |

The portable requirements.txt defaults to the 4.57.1 combination. Existing 5.19 environments remain compatible. Both combinations passed the CPU regression suite, including actual Qwen model forward/cache tests; 4.57.1 was also checked through browser canvas generation, transcription and word alignment. Other versions within the allowed range have not all been tested.

To select the tested 4.57.1 combination, run with your ComfyUI or studio Python:

python -m pip install "transformers==4.57.1" "tokenizers==0.22.2" "huggingface-hub==0.36.2"

Portable code updates preserve the installed Python packages; they do not run pip or automatically switch Transformers versions. No model downloads are needed for this compatibility update.

Canvas workflows

The workflows/ directory contains 15 canvas workflows: basic speech, references, revision, infill, long text, streamed text, scenes, SRT, codec, Qwen ASR, R2T2, alignment, live preview, session save and session load. Open a JSON file in ComfyUI and click Run. These examples were tested through the browser canvas, including visible text results and audio playback.

Select your own audio in LoadAudio; reference and alignment transcripts must match the recording. Run workflow_save_session.json before loading a session, then refresh ComfyUI and select the saved file. Long text decodes all segments. Numeric seeds support 0–9007199254740991; for larger seeds, enter the exact decimal value in seed_text (up to 18446744073709551615).

Reference-conditioned scenes assign named roles stable <|speaker:N|> IDs in first-appearance order, retained across revisions. Reference names are display labels; use matching explicit speaker tags in reference transcripts and scripts to bind particular voices.

Models

Download from t8star/Drama-3-Comfy. Default paths are relative to ComfyUI/.

| Model | Default location | |---|---| | DRAMA FP8, 38.58 GB | models/diffusion_models/drama3-t8-fp8.safetensors | | Complete text/control head, 1.03 GB | models/diffusion_models/drama3-text-head.safetensors | | Codec, 1.57 GB | models/vae/drama3-codec.safetensors | | Qwen ASR, optional | models/asr/qwen3-asr-06b.safetensors | | ForcedAligner, optional full directory | models/asr/qwen-aligner/ | | R2T2 Q8 model + projector, optional | models/asr/Confucius4-R2T2-GGUF/ | | FireRedVAD, for segmented R2T2 | models/asr/FireRedVAD-ONNX/ |

Keep all three DRAMA files for complete functionality. Each loader supports custom file/directory paths, including another drive. The studio stores these settings under userdata/.

Local studio and updates

The author distributes the complete Windows environment separately. GitHub Releases contain program files and native libraries only, without model weights or a Python runtime. Studio source is in portable/.

For an existing complete bundle based on 0.1.3 or later, close the studio and extract Drama3-T8-Portable-0.1.11-code.zip into the directory containing Start-Drama3.cmd. Double-click Start-Drama3.cmd to launch. Subsequent launches check stable GitHub Releases and apply compatible code updates; Update-Drama3.cmd checks manually while the studio is closed. Updates verify sizes and SHA-256, back up old files and roll back on failure. Models, Python, custom paths, projects, voice libraries and outputs are retained. Offline checks allow startup; dependency/runtime changes require a new complete environment from the author.

Hardware and acceleration

Validated: Windows x64, RTX 5090 Laptop 24 GB VRAM / 64 GB RAM, Torch 2.8 + CUDA 12.8, 12 GiB expert cache. FP8 MoE experts stay in system RAM; cache 0 selects an automatic budget. Leave VRAM for resident weights, codec and context. Other hardware configurations have not been validated.

Default: Torch linear attention + SDPA. Optional FLA is integrated but showed no clear speed gain on this setup. Installing FlashAttention alone does not enable it in the model; CUDA Graph is disabled. CPU previews arrive during generation, but the tested offload setup generates below real time. Transcription may require manual correction. Scene revisions preserve the full session and use bounded line-boundary recovery. Empty, extra or unfinished segments are rejected.

Credits and licenses

Based on Fish Speech, PedroMarinhoDev/ComfyUI-Drama3, ComfyUI and the linked R2T2 project. T8 additions include complete control output, stateful editing, streaming, native ASR integration and the local studio.

Project code: GPL-3.0, with upstream components retaining their licenses. DRAMA materials and weights: Fish Audio Community License v1.0; Qwen and FireRedVAD: Apache-2.0; Confucius models: NetEase Youdao Model Use License Agreement; llama.cpp: MIT. See licenses/ and upstream-lock.json. This project is not affiliated with or endorsed by upstream authors.

T8star

Bilibili · YouTube · API · Free gallery · Online AI apps · ComfyUI bundle · Hugging Face