ComfyUI Node
FireRedTTS3 Load Model
Load FireRedTTS3 (base or instruct) natively with ComfyUI/AIMDO memory registration.
FireRedTTS3 Load Model
- firered_model
◄repoFireRedTTS3-bf16►
◄variantfireredtts3_instruct►
◄dtypeauto►
◄deviceauto►
◄attentionauto►
◄download_if_missingtrue►
CategoryFireRedTTS3
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| repo | COMBO | FireRedTTS3-bf16 | Weight source. bf16 keeps the flow head/decoder in fp32 and matches official mixed-precision compute (recommended). int8 ConvRot quantizes the transformer linears (smallest, experimental). fp32 is the official full-precision repo. Missing files download when download_if_missing is on; otherwise the error tells you where to place them. |
| variant | COMBO | fireredtts3_instruct | base: zero-shot cloning in 24 languages + 21 Chinese dialects using language tags. instruct: voice design (Chinese/English only) + speech editing; its cloning inherits the language from the reference audio (no language tags). |
| dtype | COMBO | auto | bf16 stores the backbone LLM and RedAE encoder in bf16 (same compute as the official autocast path) and keeps the flow head/decoder fp32. fp32 is full precision. auto picks bf16 on supported GPUs. |
| device | COMBO | auto | Device for inference. auto follows ComfyUI's current torch device; cpu is a slow fallback. |
| attention | COMBO | auto | Attention backend for the Qwen3 transformers. auto uses flash_attention when flash_attn is installed and compatible (CUDA + bf16 compute), else sdpa. The fp32 RedAE decoder always uses sdpa. |
| download_if_missing | BOOLEAN | true | Download the selected weights, RedAE codec, tokenizer, CAM++ and FastText language-ID files into ComfyUI/models/fireredtts3 when missing. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| firered_model | FIREREDTTS3_MODEL | — |