VELVET VICE MiniMax H3 — Power LoRA AV

H3 is a rare beast among video models: because it generates video and audio together in one context, its LoRAs can be modality-scoped. A LoRA can affect the whole output, or just the video channel, or just the audio - think a style LoRA that changes the look without touching the soundtrack, or a voice/character LoRA that shouldn't disturb the visuals. VelvetViceMiniMaxH3PowerLoraAV is the node that lets you stack several of these in a defined order, each with its own strength, and apply them as a group.

The name and the concept both come from the broader Velvet Vice Power LoRA pattern, and the "AV" suffix is the H3-specific twist: FULL / VIDEO / AUDIO routing per entry. That's the "ordered MiniMax H3 LoRA stack with FULL, VIDEO and AUDIO routing" from the pack README.

How it works

The node takes your diffusion model and clip in, plus the stack definition, and outputs the patched model and clip out. The stack itself arrives as lora_stack_json (STRING) - a JSON payload that lists the LoRAs in application order with their settings - while lora_catalog is a dropdown of available LoRA files from your loras folders (defaulting to None, meaning you drive it from the JSON or the frontend catalog rather than a single picker).

Order matters here in a way it doesn't for a single LoRA. LoRA application isn't commutative in general: A then B can land differently from B then A, and when you're mixing a full-model style LoRA with a video-only motion LoRA and an audio-only LoRA, the sequence and per-LoRA strength are exactly what you're trying to control. The pack's warning is worth quoting: "Use H3/FL2VA-compatible LoRAs only." This node will happily apply a LoRA that was trained for a completely different model family and produce garbage - the compatibility gate is on you, not the node.

Inputs and outputs

• model (MODEL), clip (CLIP) - in.
• lora_stack_json (STRING) - the ordered stack, as JSON.
• lora_catalog - LoRA file picker (None default).

Outputs: model (MODEL), clip (CLIP).

Install and practical notes

Same pack, same install:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3

then restart. The LoRAs themselves are downloads - H3 Turbo LoRAs are linked from the README on Civitai, and FL2VA-compatible fine-tune LoRAs live in the usual places.

Where beginners get confused: they expect to load a stack of regular SD LoRAs here because the node is called "Power." It isn't a generic LoRA powerhouse - it's H3/FL2VA-specific, and the AV routing is only meaningful if your LoRAs actually declare a modality. If your LoRA isn't H3-compatible, use ComfyUI's stock LoRA loader for it. And if a stack applies but the audio channel is unchanged while video shifts, that's the routing working as intended, not a bug - check which entries you marked VIDEO-only.

Also note the pack draws a bright line between this and the separate single-purpose Turbo loader: OptionalTurboLora handles the dedicated turbo/distilled LoRA on its own switch, and PowerLoraAV is for general ordered stacks. Wiring the turbo LoRA into the stack and the turbo path doubles it up - keep them in their lanes.