ComfyUI Node
Raon OpenTTS Load Model
Load Raon-OpenTTS (0.3B or 1B) natively with ComfyUI/AIMDO memory registration.
Raon OpenTTS Load Model
- raon_model
◄weightsRaon-OpenTTS-1B-int8-convrot.safetensors►
◄dtypeauto►
◄deviceauto►
◄attentionauto►
◄download_if_missingtrue►
CategoryRaonOpenTTS
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| weights | COMBO | Raon-OpenTTS-1B-int8-convrot.safetensors | Checkpoint file from ComfyUI/models/raon_opentts (0.3B and 1B in separate folders). int8-convrot is the quantized build (smallest, runs through comfy-kitchen INT8 kernels); bf16 is the half-precision reference. The fp32 extraction is conversion reference only and is not listed. With download_if_missing on, only the selected build (plus its config/vocab and the vocoder) is downloaded — never the whole repo. |
| dtype | COMBO | auto | Compute/storage dtype for the DiT. auto picks bf16 on supported GPUs. INT8 weights always stay INT8 and their scales fp32 regardless of this choice. |
| 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 DiT blocks. auto uses flash_attn (FA2) when installed with half-precision compute, otherwise torch SDPA. sageattention patches SDPA at generation time. |
| download_if_missing | BOOLEAN | true | Download the HiFi-GAN vocoder into ComfyUI/models/raon_opentts when missing. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| raon_model | RAON_OPENTTS_MODEL | — |