ComfyUI-Drama3: Fish Audio Drama 3 Runs Locally on 8 GB Cards

ComfyUI-Drama3: Drama 3 and Qwen3-ASR as ComfyUI nodes, no extra packages to install.

Drama 3 is a 35B mixture-of-experts text-to-speech model with 3B parameters active per token. Fish Audio's own package is eight safetensors shards plus a codec, roughly 40 GB on disk, and the ComfyUI partner node reaches it over the API. This pack is the local route: it loads the released checkpoints itself, generates audio codes inside ComfyUI, and decodes them to 44.1 kHz on the same graph.

A 35B model on an 8 GB card

Drama 3 is a Qwen3.5-style mixture of experts: 40 layers, each with 256 routed experts of which 8 run per audio frame, plus attention or DeltaNet, a shared expert and a 13-layer codebook decoder. Everything except the routed experts is about 6 GB and stays on the GPU. With experts in RAM, the 32 GB of expert weights sit in pinned system memory: each frame copies the eight experts per layer it needs, and a least-recently-used cache in the VRAM that is left free keeps the ones that come back, with 85 to 95 percent of reads hitting a 13 GB cache.

The pack ships four builds of the checkpoint and lets you pick per machine:

FP8 keeps Fish Audio's own expert weights byte for byte. int8 is the same size and ran 14 to 17 percent faster on an RTX 3090, because int8 experts run as int8 while FP8 experts are turned into bf16 on every pass. Smaller files are for machines short of the FP8 file's ~48 GB of RAM. A second GPU can take the expert weights when experts is set to VRAM instead of RAM.

Voices, dialogue and cues

The pack keeps to Drama 3's own prompt syntax, the same one the technical report documents:

Up to 16 speakers can share one text, and each one gets a voice slot. Long text is cut into chunks of about 300 bytes without splitting a tag. Take generation has four modes: one take per line, one take per line that hears the line before, one take per speaker, or the whole script together so every line hears the others. Decode joins separate takes with 0.3 seconds of silence and can shorten long pauses.

A voice is encoded once from a clip plus its transcript and can then be saved as a few-kilobyte file in output/drama3_voices, so later runs need neither the clip nor the codec. Generated takes can be saved the same way in output/drama3_codes and re-decoded or edited without generating again.

Transcribing and editing takes

Six samples from the pack's samples page: cloning, cues, a two-speaker scene, a five-speaker scene, singing and Portuguese. Each was generated on an RTX 3090 or RTX 4090 with the FP8 build.

Voice cloning starts with a transcript, and the pack supplies one: Load Qwen3-ASR plus Qwen3-ASR Transcribe run a 1.7B or 0.6B speech recognition model, and a forced aligner, in 30 languages with no extra packages. On an RTX 3090 a 6.6 second clip transcribes in about 2 seconds with the 1.7B model and about a second with the 0.6B, model load included. The same aligner feeds Drama 3 Editor, which opens over the ComfyUI canvas: captions follow the voice word by word, one track per voice, and the editor cuts, adds silence, names voices and sets their volume. Pressing Use these edits runs the node again with them, and they stay in the workflow.

Speed

Drama 3 makes 21.5 audio frames per second of speech, so 46 ms per frame is real time. Decoding one frame at a time is CPU-bound, since a single core launches many small GPU kernels, so the expert cache size moves the speed less than expected. Measured with a cloned voice and about 9 seconds of speech on an Intel Core i9-14900K, prompt pass included:

The last two rows ran with the rest of the 3090's memory taken by another process, standing in for a 12 GB or 8 GB card. The author also warns that Fish Audio's server sends top_k 30, which makes takes go silent at pauses, so the nodes default to the Space's own settings of temperature 1.0, top_p 0.9 and no top_k cap.

Getting started

cd ComfyUI/custom_nodes
git clone https://github.com/PedroMarinhoDev/ComfyUI-Drama3

hf download PedroMarinhoDev/Drama-3-ComfyUI drama3_fp8experts_bf16.safetensors --local-dir ComfyUI/models/diffusion_models
hf download PedroMarinhoDev/Drama-3-ComfyUI drama3_codec.safetensors --local-dir ComfyUI/models/vae
hf download PedroMarinhoDev/Qwen3-ASR-ComfyUI qwen3_asr_0.6b_bf16.safetensors --local-dir ComfyUI/models/asr

Availability

Node pack: PedroMarinhoDev/ComfyUI-Drama3
Weights: PedroMarinhoDev/Drama-3-ComfyUI · PedroMarinhoDev/Qwen3-ASR-ComfyUI
Samples: pedromarinhodev.github.io/ComfyUI-Drama3
Base model: fishaudio/drama-3 · the Drama 3 news