memory-leak/drama3

Drama 3 — INT8 ConvRot for ComfyUI

Built with Fish Audio.

An unofficial, ComfyUI-compatible conversion of Fish Audio Drama 3. Use it with ComfyUI-Drama3. This is a converted checkpoint, not a fine-tune, and is not an official Fish Audio release.

File	ComfyUI folder	Contents
drama3_int8convrot.safetensors	models/diffusion_models	INT8 ConvRot routed experts and supported dense linears; 36.34 GB
drama3_codec.safetensors	models/vae	Required audio codec; 1.57 GB

Embeddings, heads, norms, routing, and selected small projections retain their original precision. The original block-FP8 expert weights were dequantized before INT8 ConvRot quantization.

Use

1. Install ComfyUI v0.40.0 or newer and the ComfyUI-Drama3 custom nodes.
2. Download the two files into the folders above, then restart ComfyUI.
3. Open a workflow from the node pack's example_workflows folder.
4. In Load Drama 3, select drama3_int8convrot.safetensors and set experts to in RAM.
5. In Load Drama 3 Codec, select drama3_codec.safetensors.
6. For voice cloning, supply a clean reference recording and its exact transcript.

The full model does not fit into 16 GB VRAM. Use experts in RAM and a VRAM expert cache. Japanese speech generation was tested with ComfyUI v0.40.0, an RTX 5060 Ti 16 GB, 64 GB system RAM, a 4 GiB expert cache, and --disable-pinned-memory. An available amount of system RAM comparable to the model size is needed; loading can use more. No-reference generation can stop early; a reference voice generally gives more reliable results.

Conversion

Source checkpoint revision: e0a9218aa48888270f06ef1ed9ea67622c0e17d2.

Converted using ComfyUI-Drama3 revision 0465897ec86c0f655af12bf8ae698afd89a58723:

python tools/convert.py model /path/to/drama-3 drama3_int8convrot.safetensors --format int8_convrot --dense int8
python tools/convert.py codec /path/to/drama-3/codec.pth drama3_codec.safetensors

License

The weights and derived materials remain subject to the Fish Audio Community License. See Notice for attribution and modifications. Conversion does not remove the original license conditions. The custom-node implementation has its own licenses; consult its repository.