H3 Speech Voice Clone (Ref2VA)

Tiny-canvas Ref2VA voice clone. Gemmy audio-h3 speech with a reference wav.

Unofficial project. Not affiliated with or endorsed by Comfy Org.

Purpose

Clone a speaker from reference audio + dark picture.

AUTHORING SOURCE: ir.build.ts. GENERATED: workflow.ir.json, comfy.workflow.json, prompt.template.json — do not hand-edit. Rebuild with pnpm comfy:build from comfy-workflows/. Gemmy runtime binds prompt.template.json (Python, no Node).

Install

pnpm add @stepupgaming/comfy-workflow-h3-speech-ref2va

Required parameters

• prompt
• image
• reference_audio

Optional parameters

• unet (default "minimax_h3_ref2va_pruned_int8_convrot.safetensors")
• clip (default "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors")
• video_vae (default "minimax_h3_video_vae_fp16.safetensors")
• audio_vae (default "minimax_h3_audio_vae_fp32.safetensors")
• width (default 32)
• height (default 32)
• length (default 90)
• seed (default 42)
• steps (default 20)
• output_prefix (default "h3/speech_ref")

Custom node packs

• (core ComfyUI classes only, or mapping unknown)

Node classes

• BasicGuider
• BasicScheduler
• CLIPLoader
• KSamplerSelect
• LoadAudio
• LoadImage
• MiniMaxH3ReferenceToVideo
• RandomNoise
• SamplerCustomAdvanced
• SaveAudio
• UNETLoader
• VAEDecodeAudio
• VAELoader

Models

• model: minimax_h3_video_vae_fp16.safetensors
• model: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
• model: minimax_h3_audio_vae_fp32.safetensors
• model: minimax_h3_ref2va_pruned_int8_convrot.safetensors

Do not publish model binaries. Filenames are defaults from Gemmy's H3/Krea layout; they are not universally available.

Inspect & run

cwf inspect @stepupgaming/comfy-workflow-h3-speech-ref2va --url http://127.0.0.1:8188
cwf run @stepupgaming/comfy-workflow-h3-speech-ref2va --url http://127.0.0.1:8188 --param prompt=...

License

MIT for this package's Graph IR / scaffolding, owned by Gemmy. Models and third-party custom nodes remain under their own licenses.

Never hand-edit generated workflow JSON.