Dhrog/Audio8-TTS-Realistic-0.6b-ComfyUI-Packs

Audio8-TTS-Realistic-0.6b — ComfyUI model packs (BF16 + INT8)

Single-file model packs of Edge0/Audio8-TTS-Realistic-0.6b for use inside ComfyUI. Each pack folds the original model directory (language model, 44.1 kHz codec, tokenizer and processor files) into one .safetensors file that ComfyUI's diffusion_models picker can list and load.

This is a community re-packaging, not an official Edge0 release. The weights are unchanged from upstream apart from the INT8 weight quantisation described below.

ComfyUI nodes: Pepehoschi/ComfyUI-Audio8TTS (public GitHub repository, includes example workflows and setup instructions).

Files

File	Size	Contents
Audio8-TTS-Realistic-0.6b-BF16.safetensors	2,567,214,504 bytes	Original BF16 weights, packed lossless. SHA256 d9720afccadca7f162bc780775b440d5f1bbc98bf1175e62281e31c75d410e42
Audio8-TTS-Realistic-0.6b-int8convrot.safetensors	2,147,558,576 bytes	ConvRot INT8 weight quantisation (8-bit weights + per-channel scales, BF16 activations), 16% smaller. Metadata audio8_quant=int8_convrot. SHA256 cc13f4b0b6e1cd2712bb1339819988574e7e9c3ab1fa8bdc21cbddc7202e3306

Packing is lossless and deterministic: repacking the same source directory reproduces identical tensor bytes.

Usage in ComfyUI

1. Install the ComfyUI-Audio8TTS custom nodes (git clone into ComfyUI/custom_nodes/).
2. Put one of the packs into ComfyUI/models/diffusion_models/audio8/.
3. Add an Audio8 Model Loader node and pick the file from its dropdown. The loader selects the compute path from the pack metadata (audio8_quant=int8_convrot), not from the file name; dtype="auto" is the recommended setting.
4. Build a graph with the other Audio8 nodes (Audio8 TTS, Audio8 Save Voice, Audio8 Load Voice, Audio8 Encode Reference, …). Two ready-made templates ship under the ComfyUI template browser (category ComfyUI-Audio8TTS).

The original unpacked Edge0 directory keeps working as an inference source; these packs only provide the native single-file ComfyUI loading path.

Measured behaviour

Local measurements on a ComfyUI cu128 instance (256 frames, greedy sampling, cuda_graphs=true, cuda_graph_scope=frame, comfy_ops=false):

• BF16 pack: ~54 tok/s.
• INT8 pack: 30.4 tok/s (eager 32.62 s, fast 16.44 s, frame 8.43 s for 256 frames). BF16 remains the faster path; INT8 mainly saves memory and download size.
• INT8 sampling is bit-identical across eager / fast / frame modes for the same pack, and the INT8 pack is bit-identical to on-the-fly INT8 quantisation of the BF16 pack (maxdiff 0.0 over a 491,520-sample render).
• The INT8 quantisation itself is deterministic across fresh loads.

Provenance and licence

• Base model and architecture: Edge0/Audio8-TTS-Realistic-0.6b, Apache-2.0. The upstream LICENSE and NOTICE files are included in this repository unchanged. These packs were built from that repository (recorded in each pack's audio8_model_id metadata). Note: on 2026-11-02 Edge0 also published the same release under the new name Edge0/Audio8-TTS-Realistic-Air; both repositories carry byte-identical model.safetensors and codec.pth LFS objects, so either upstream matches these packs. The predecessor Edge0/Audio8-TTS-Preview-0.6b shares only the codec, not the language model weights, and is not the source of these packs.
• ComfyUI node code, packing tooling and tests: Pepehoschi/ComfyUI-Audio8TTS.
• These packs are therefore distributed under Apache-2.0 as well.

Limitations

• The upstream model supports 23 languages; per-language quality varies — see the upstream model card for evaluations.
• INT8 is weight-only quantisation; activations stay BF16. No accuracy claims beyond the measurements above are made.
• The packs target the ComfyUI node workflow; plain Transformers inference should use the original upstream repository.