MiniMax H3 REF2VA Turbo LoRA (CRT AutoDL)

MiniMax H3's claim to fame isn't just pretty video - it's that the picture and the sound come out of the same model, in one pass. The REF2VA mode is the omni-modal side of that: you feed it reference images, videos and audio, and it generates a clip conditioned on all of them, audio included. The catch is that full-quality sampling is slow. This node applies the REF2VA Turbo LoRA - the 4-step distillation for that mode - and, being a CRT AutoDL node, it downloads the LoRA itself on first use.

This is the sibling of the FL2VA Turbo LoRA node, and the distinction matters. FL2VA is the checkpoint family for T2V and first/last-frame I2V; REF2VA is the reference-to-video family - up to nine reference images, three reference videos and three standalone audio clips in one generation (twelve files mixed is the official cap). Different family, different file. If you're doing prompt-only or keyframe work, you want the FL2VA node, not this one.

How it works

Same fixed-loader pattern as the rest of the AutoDL family. On first run it looks in ComfyUI/models/loras/ for the file, downloads it from lightx2v/Minimax-h3-Turbo if missing (progress bar in the console, prefixed CRT AutoDL), and applies it with ComfyUI's standard LoRA loader. The file is minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors - 4-step, bf16. Note the version: v0.1 for REF2VA versus v1.1 for the FL2VA LoRA. The reference mode's turbo LoRA is the younger, less refined of the two, which tracks - R2V with audio conditioning is the harder problem.

Inputs and output

Two inputs, that's it:

• model - the REF2VA diffusion model, typically from the pack's MiniMax H3 Model (CRT AutoDL) node or another REF2VA loader.
• strength_model - default 1.0, range -100 to 100.

Output is a single MODEL.

Where it slots in: in the CRT unified-sampler setup you feed it to the MiniMax H3 US Models Pipe node's ref2va_turbo_model socket and let the sampler's Turbo toggle handle the rest. Standalone, apply it to your base model and sample at around 4 steps.

Install

This is one node in CRT-Nodes by PGCRT. ComfyUI Manager → search CRT-Nodes → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

Restart ComfyUI after. If an update leaves a node red or showing NaN, right-click → Fix node (recreate).

Gotchas

Four steps is what the LoRA was trained for - don't be tempted to crank it. Strength 1.0 is the setting to start from. And the usual H3 housekeeping applies: the pack's requirements.txt is a heavy list (whisper, librosa, ultralytics and friends) even though the H3 nodes don't need most of it. One more thing, because it's easy to miss with any H3 content: the H3 Community License geofences the local weights out of the US, EU, UK and South Korea - outputs included - so check your side of that line before you build a workflow on it.