PedroMarinhoDev/Drama-3-ComfyUI

Nodes on GitHub · All samples · Qwen3-ASR for the transcripts · The original model

Fish Audio Drama 3, converted for the native ComfyUI nodes of ComfyUI-Drama3: a 35B text-to-speech model for voice cloning, dialogue and inline cues such as [sighs] or [whispers], running on cards from 8 GB. 90% of the model is routed experts and each audio frame uses only 8 of the 256 in a layer, so the nodes keep the experts in system RAM and a cache of the ones in use on the GPU. Each file holds the whole model in one file; the codec is separate. Built with Fish Audio.

Samples

No real person is cloned here: the first three clone voices the model made up from a single line, the rest are voices it invents from the prompt alone. Drama 3 FP8, experts in RAM, temperature 1.0, top_p 0.9, top_k 30 (the nodes now default to top_k 0, as Fish Audio's Space does: top_k 30 made some takes fall silent and drop a sentence). More on the demo page →

Voice cloning: the source clip, a voice the model made up from one line, then a new line in that voice

The lighthouse keeper climbed the last step, set her lantern on the sill, and watched the storm roll in across the bay. Somewhere below, a bell began to ring.

Inline cues

[whispers] Don't move. I think it heard us. [pause] [sighs] Oh, it's only the cat. [laughs] You should have seen your face.

Two speakers

<|speaker:0|> Did you remember the tickets?
<|speaker:1|> [laughs] I thought you had them!
<|speaker:0|> [sighs] Then we're watching the concert from the parking lot.

Five speakers, no voices given: the model casts all five

<|speaker:0|> [dramatic dungeon master] You enter the dark cave. A dragon sleeps on a pile of gold.
<|speaker:1|> [brave, too loud] I attack the dragon! …

Singing

[singing, grand operatic] Oh toast, my golden toast, you're the bread I love the most! [singing softly] Crispy edges, warm and brown, finest slice in all the town. …

Portuguese

[Brazilian football fan, building tension, Portuguese] Vai, vai, vai, vai! Ele passou por um, passou por dois, [shouting] olha o chute! …

Files

Start with FP8: it holds Fish Audio's own expert weights, byte for byte, next to the checkpoint's bf16 layers. The int8 file is the same size, measured as close to the original, and ran 14–17% faster than FP8 on an RTX 3090: the nodes turn FP8 experts into bf16 each time they run, while int8 experts run as int8. Take a smaller file when your system RAM is short of the FP8 file's ~48 GB. Every model file needs the codec.

File	Size	Experts	System RAM	Notes
drama3_fp8experts_bf16.safetensors	38.6 GB	FP8, as Fish Audio ships them	~48 GB	the original weights
drama3_int8experts_bf16.safetensors	38.7 GB	int8 ConvRot	~48 GB	faster than FP8 on the 3090; no silent take in 12 cloned test takes
drama3_w6a8experts_bf16.safetensors	32.7 GB	W6A8	~40 GB	no silent take in 16 cloned test takes
drama3_w4a8experts_bf16.safetensors	24.6 GB	W4A8	~32 GB	falls silent more often with real voice clips
drama3_codec.safetensors	1.6 GB			the original codec, fp32

Quick start

1. Install the nodes: clone ComfyUI-Drama3 into ComfyUI/custom_nodes and restart ComfyUI. Nothing to pip install.

2. Download a model file into models/diffusion_models and drama3_codec.safetensors into models/vae. For transcripts, add qwen3_asr_1.7b_bf16.safetensors from Qwen3-ASR-ComfyUI to models/asr, and for the Drama 3 Editor's word-by-word captions, qwen3_forced_aligner_0.6b_bf16.safetensors from the same repository.

   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_1.7b_bf16.safetensors --local-dir ComfyUI/models/asr
   hf download PedroMarinhoDev/Qwen3-ASR-ComfyUI qwen3_forced_aligner_0.6b_bf16.safetensors --local-dir ComfyUI/models/asr
   

3. Open the workflow example_workflows/drama3_voice_clone.json from the pack, pick a few seconds of clean speech, write your text, run.

4. Edit the take (optional): the Drama 3 Editor node, in place of Decode, shows the take on a timeline with one track per voice and captions that follow the words; it shortens long pauses, spaces out lines that run into each other, flags lines that may have lost words, and cuts. For dialogue, give each speaker a voice: the voice in slot voiceN speaks for <|speaker:N|>.

Requirements and speed

GPU	NVIDIA, 8 GB of VRAM or more with experts in RAM (tested: RTX 4090 and RTX 3090; 12 GB and 8 GB simulated on the 3090)
System RAM	the model file plus ~8 GB
ComfyUI	v0.40 or newer

Drama 3 makes 21.5 audio frames per second of speech, so 46 ms per frame is real time. Experts in RAM, a cloned voice, ~9 s of speech, prompt pass included, measured on 2026-10-15 with an Intel Core i9-14900K (the held rows on 2026-10-14). One frame at a time, decoding waits on the CPU: int8 runs at the same 69 ms per frame on both GPUs, and a CPU with slower cores adds to these times:

GPU	File	Expert cache	ms per frame
RTX 4090 (PCIe 4.0 ×8)	W4A8	16 GB	52
RTX 4090 (PCIe 4.0 ×8)	int8	16 GB	69
RTX 3090 (PCIe 4.0 ×8)	int8	16 GB	69
RTX 3090 (PCIe 4.0 ×8)	FP8	16 GB	84
RTX 3090 held to 12 GB	FP8	automatic	137
RTX 3090 held to 8 GB	FP8	automatic	214

How these files were made

From fishaudio/drama-3 with the pack's tools/convert.py:

• FP8: the routed experts' FP8 bytes and 128×128 block scales copied unchanged, gate and up projections fused.
• int8 / W6A8 / W4A8: the experts dequantized from FP8 in fp32 and requantized, in ComfyUI's comfy_kitchen formats: int8 ConvRot (per-channel int8 after a rotation in groups of 256), or 6-bit or 4-bit codebook weights with int8 activations; relative weight error 0.83%, 2.1% and 7.3%.
• Every file: the attention, DeltaNet, shared-expert and codebook-decoder layers are the checkpoint's own bf16; the LM head keeps only the rows generation samples from (end of speech and the 4096 semantic codes).
• Codec: Fish Audio's S2-Pro codec as shipped with Drama 3, weight norm folded in, fp32.

Against Fish Audio's own PyTorch code, with the same frames forced, the main token's distribution differs by a KL of 0.004–0.008 nats with the FP8 and int8 files, the noise of bf16 arithmetic.

License

These Fish Audio Materials are licensed under the Fish Audio Community License, Version 1.0. Copyright © 39 AI, INC. All Rights Reserved. See LICENSE and NOTICE.

Free for research and non-commercial use, and for commercial use below USD 500k annual revenue; a hosted TTS service needs a separate license from Fish Audio. Don't use these files or their output to imitate the voice of an identifiable person without their consent, or to train or improve other speech models.

Built with Fish Audio. This repository is not affiliated with, sponsored or endorsed by Fish Audio.