ComfyUI Node
LongCat AudioDiT TTS
LongCat-AudioDiT Text-to-Speech. Zero-shot TTS with diffusion-based waveform generation.
LongCat AudioDiT TTS
- audio
◄model_path▾►
◄textHello! This is your Longcat Audio node speaking, everything is set-up and running smoothly!►
◄steps16►
◄guidance_strength4.0►
◄guidance_methodcfg►
◄deviceauto►
◄dtypeauto►
◄attentionauto►
◄seed0►
◄keep_model_loadedtrue►
CategoryLongCat-AudioDiT
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| model_path | COMBO | LongCat-AudioDiT model. Models are stored in ComfyUI/models/audiodit/ | |
| text | STRING | Hello! This is your Longcat Audio node speaking, everything is set-up and running smoothly! | Text to synthesize. |
| steps | INT | 164–64 | Number of ODE Euler steps. More steps = better quality but slower. |
| guidance_strength | FLOAT | 4.00–10 | CFG/APG guidance strength. Higher = more guidance. |
| guidance_method | COMBO | cfg | Guidance method. 'apg' often gives better results for voice cloning. |
| device | COMBO | auto | Compute device. 'auto' picks CUDA > MPS > CPU. |
| dtype | COMBO | auto | Model dtype. 'auto' picks bf16 for CUDA, fp16 for MPS, fp32 for CPU. |
| attention | COMBO | auto | Attention implementation. 'auto' uses model default (SDPA). 'sage_attention' requires sageattention package. 'flash_attention' forces FlashAttention via SDPBackend. |
| seed | INT | 00–2147483647 | Random seed. 0 = random. |
| keep_model_loaded | BOOLEAN | true | Keep model loaded between runs. Model is automatically offloaded to CPU after generation to free VRAM, then resumed to GPU on the next run. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| audio | AUDIO | — |