t8star/Drama-3-Comfy

Drama 3 · T8 ComfyUI models

English | 简体中文

Built with Fish Audio.

English

Converted DRAMA-3 weights for ComfyUI-Drama3-T8 and its local studio. The FP8 format keeps MoE experts in system RAM and preserves the complete text/control output head for scenes, revisions, infill and streaming. These files require the T8 loader; they are not a Transformers checkpoint.

File / directory	Size	ComfyUI destination
drama3-t8-fp8.safetensors	38.58 GB	models/diffusion_models/
drama3-text-head.safetensors	1.03 GB	models/diffusion_models/
drama3-codec.safetensors	1.57 GB	models/vae/
qwen3-asr-06b.safetensors	1.56 GB	models/asr/
qwen-aligner/	1.84 GB	models/asr/qwen-aligner/
Confucius4-R2T2-GGUF/	2.18 GB	models/asr/Confucius4-R2T2-GGUF/
FireRedVAD-ONNX/	2.3 MB	models/asr/FireRedVAD-ONNX/

The first three files provide complete DRAMA synthesis. ASR, alignment and VAD are optional per workflow. Keep the R2T2 Q8 main model and matching projector together. Download complete folders for the aligner and VAD. All paths can be overridden in the nodes or studio settings; the studio uses the same layout directly under its models/ directory.

Validated on Windows x64, RTX 5090 Laptop 24 GB VRAM / 64 GB RAM, Torch 2.8 + CUDA 12.8, with a 12 GiB expert cache. This offload setup generates below real time. R2T2 requires the separately distributed Windows CUDA native runtime. See the code repository for installation and supported acceleration; weights alone do not enable FlashAttention. Use code v0.1.10 or later for scene-revision boundary recovery, stable reference speaker casting and strict completion checks. Model weights are unchanged.

Code v0.1.11 supports Transformers 4.57.1 with Tokenizers 0.22.2 and Hugging Face Hub 0.36.2; existing 5.19 environments remain compatible. Generation, transcription and word alignment were checked on 4.57.1 through actual ComfyUI canvas runs. See the code README for installation; model weights are unchanged.

DRAMA files derive from fishaudio/drama-3, revision e0a9218aa48888270f06ef1ed9ea67622c0e17d2, under the Fish Audio Community License v1.0 in LICENSE. Qwen ASR/aligner use Apache-2.0; Confucius uses the NetEase Youdao Model Use License Agreement; FireRedVAD uses Apache-2.0. Full licenses are in licenses/, fixed sources in upstream-lock.json, and file sizes / available checksums in model-manifest.json. Auxiliary models retain their own licenses. This redistribution is not affiliated with or endorsed by the upstream authors.

简体中文

适用于 ComfyUI-Drama3-T8 与本地演出工作台的转换模型。FP8 主模型将 MoE 专家保留在系统内存，完整文本输出头保留场景、导演修改、局部编辑及流式功能所需控制信息。请使用 T8 加载器，不可直接作为 Transformers 模型加载。

前三个文件为完整语音生成功能所需；转写、对齐与 VAD 按工作流选用。ComfyUI 放置路径见上表。R2T2 的主模型和投影器必须配套；对齐及 VAD 请下载完整目录。所有模型均支持自定义路径；整合包对应放在根目录的 models/ 下。

实测环境：Windows x64、RTX 5090 Laptop 24 GB 显存、64 GB 内存，Torch 2.8 / CUDA 12.8，专家缓存 12 GiB。该卸载配置低于音频实时生成速度。R2T2 还需要代码仓库单独分发的 Windows CUDA 原生运行库。加速支持以代码仓库说明为准。场景改稿请使用 v0.1.10 或后续代码，包含边界恢复、稳定的参考角色映射与完整性校验；模型权重未变。

代码 v0.1.11 支持 Transformers 4.57.1、Tokenizers 0.22.2、Hugging Face Hub 0.36.2 这组依赖；已有 5.19 环境也可继续使用。4.57.1 已经实际 ComfyUI 画布生成、转写及词级对齐验证。安装方式见代码仓库 README，模型权重未变。

DRAMA 权重遵循 Fish Audio Community License v1.0，Qwen 转写/对齐为 Apache-2.0，Confucius 遵循网易有道模型使用许可，FireRedVAD 为 Apache-2.0；辅助模型分别保留其原有许可。完整文本、来源固定版本及文件清单见上方链接。

T8star

Bilibili · YouTube · API · Free gallery / 免费画廊 · Online AI / 在线AI应用 · Portable bundle / ComfyUI整合包 · Hugging Face