ComfyUI Node Runs on cloud
H3 Sound LoRA
Put a LoRA on MiniMax H3's sound tokens alone, so the picture keeps the model's own LoRAs, such as Pika H3 Flex, while the sound takes another, such as DMAD. Sets the audio shift that LoRA was trained at. Chain a second one on add to mix two on the sound.
H3 Sound LoRA
- model
- model
◄lora_nameNone►
◄strength1.00►
◄modereplace►
◄audio_shift2.0►
CategoryWAS Suite/Sampling
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | MiniMax H3 after every LoRA loader, Pika H3 Flex LoRA included; the picture keeps those LoRAs. | |
| lora_name | COMBO | None | LoRA in models/loras for the sound tokens, as `dmad_minimax_h3_4step_full_critic_comfyui`. Its text refiner pairs are not used. A wired name, or a full path inside that model's folders such as `E:\models\loras\style.safetensors`, works too. |
| strength | FLOAT | 1.00-2–2 | Sound LoRA scale: `1.0` = as trained; `0` = none, the model's LoRAs still dropped on replace. |
| mode | COMBO | replace | `replace` = the sound tokens drop the model's other LoRAs and take this one; `add` = this one on top of them, as a second sound LoRA in a chain. |
| audio_shift | FLOAT | 2.00.5–24 | Audio flow shift the sound samples at: `2` = DMAD; `3` = Pika H3 Flex and Ref2VA. The video shift is kept. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | The model with its sound tokens on this LoRA at this audio shift. H3 Stream Sampler steps the sound on that shift's sigmas. |