Comfy-MIMOASR

Comfy-MIMOASR wraps Xiaomi MiMo-V2.5-ASR as native ComfyUI audio nodes.

Nodes

Model Layout

By default the loader uses ComfyUI's main models folder:

ComfyUI/
  custom_nodes/
    Comfy-MIMOASR/
      vad/
        silero_vad.jit
  models/
    MiMo-Audio-Tokenizer/
    MiMo-V2.5-ASR/
    MiMo-V2.5-ASR-quantized-cache/

If either model folder is missing, the loader can download from Hugging Face or hf-mirror:

• XiaomiMiMo/MiMo-V2.5-ASR
• XiaomiMiMo/MiMo-Audio-Tokenizer

Set download_source to offline if you want missing models to fail instead of downloading.

Install

Place this folder under ComfyUI/custom_nodes/Comfy-MIMOASR, then install the requirements in the same Python environment used by ComfyUI:

pip install -r requirements.txt

Restart ComfyUI after installation.

Usage

1. Add MiMo ASR Load Model.
2. Connect its mimo_asr_model output to MiMo ASR Audio To Text.
3. Connect a ComfyUI native AUDIO output to the audio input.
4. Use the returned STRING output as transcript text.
5. Optionally connect the transcript to MiMo ASR Save Transcript to write a transcript file.
6. Optionally run MiMo ASR Release Model when you want to unload weights.

The loader exposes the same core loading controls as the original demo: device placement, mel/audio-tokenizer device, quantization, quantization cache, chunk size, progress logging, and bitsandbytes-compatible options.

The transcription node exposes language hint, generation token limit, audio start offset, duration clipping, and batch selection for batched ComfyUI audio. The default generation token limit is 128.

For timestamped transcription:

1. Connect ComfyUI native AUDIO to SILERO-VAD Audio Segmenter.
2. Keep vad_model_path as auto when vad/silero_vad.jit exists in this extension folder.
3. Connect timed_audio to MiMo ASR Timed Audio To Text.
4. Choose bracket, srt, or jsonl timestamp output. Use jsonl when you want the save node to generate .srt or .ass subtitles.

The timed audio output is a dictionary with a segments list. Each segment contains start, end, duration, sample coordinates, and a ComfyUI-style audio dictionary for that slice.

The save transcript node defaults to ComfyUI's output directory and names files like transcript_00001_.txt. Use filename_prefix for output subfolders or output_directory for a custom save directory. file_format controls the extension and content conversion. srt and ass are generated from timestamped JSONL lines with start, end, and text fields, such as the jsonl output from MiMo ASR Timed Audio To Text.

Long-running MiMo ASR nodes check ComfyUI's interrupt flag during VAD scanning, audio preprocessing, segment loops, and generation, so the ComfyUI cancel button can stop queued work instead of waiting for the full transcription to finish.

Localization

Simplified Chinese node definitions are included under locales/zh/nodeDefs.json with a locales/zh-CN/nodeDefs.json compatibility copy. They translate node names, input and output labels, tooltips, and combo options in ComfyUI's language system.

Documentation

• Original model README: doc/mimo-v2.5-asr.md
• Original demo/API notes: doc/demo.md