MiniMax H3 FL2VA Turbo LoRA (CRT AutoDL)

MiniMax H3 is genuinely good - a 33B omni-modal video model with native audio - but full-quality sampling is a 20+ step affair. This node is the shortcut: it applies the FL2VA Turbo LoRA, the 4-step distillation for the FL2VA checkpoint family, and because it's a CRT "AutoDL" node it downloads the LoRA itself the first time you run it. No hunting through HuggingFace, no mismatched filenames.

First, "FL2VA" - worth unpacking because the model naming is the thing that trips people. FL2VA is the checkpoint family H3 uses for both text-to-video (T2V) and first/last-frame image-to-video (I2V). If your clip starts from a prompt alone, or is generated between two keyframes, you're on FL2VA, and this is the turbo LoRA for you. The sibling node, MiniMax H3 REF2VA Turbo LoRA, covers reference-to-video instead - different family, different file, don't cross-wire them.

How it works

Under the hood it's a fixed LoRA loader, not a sampler. On first run the node checks ComfyUI/models/loras/ for the file, downloads it from the lightx2v/Minimax-h3-Turbo repo if it's missing (watch the console for the CRT AutoDL progress bar), and applies it with ComfyUI's standard LoRA machinery. The filename tells you everything: minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_bf16.safetensors - trained for 4 steps, at 768p, bf16.

Inputs and output

Just two inputs, which is the whole point:

• model - the FL2VA diffusion model, typically from the pack's MiniMax H3 Model (CRT AutoDL) node or any other FL2VA loader.
• strength_model - default 1.0, range -100 to 100. At 1.0 you're applying the LoRA as trained; you can dial it down for a milder turbo effect, though with distilled LoRAs you usually want the full strength.

Output is a single MODEL, ready to feed a sampler.

Two ways to use it. In the CRT unified-sampler ecosystem, wire it into the MiniMax H3 US Models Pipe node's fl2va_turbo_model socket, then flip the sampler's Turbo toggle and set steps_turbo to 4. Standalone, load the base model, apply this LoRA, and set your KSampler to about 4 steps - that's where this thing sings.

Install

This node ships in CRT-Nodes by PGCRT. Easiest: ComfyUI Manager → search CRT-Nodes → install → restart. Manual:

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

Restart ComfyUI afterwards. If you update the pack and a node shows red or NaN widgets, right-click it and pick Fix node (recreate) - the author changes sockets between versions.

Gotchas

The LoRA is trained for 4 steps and 768p. Running it at 20 steps isn't "more quality," it's just slower with a distilled LoRA that wasn't built for it. Strength 1.0 is the norm. And note the AutoDL family has a couple of heavyweight friends - the pack's requirements.txt pulls in things like librosa, whisper and ultralytics whether you use them or not.

One honest flag before you go: MiniMax H3's local weights ship under the H3 Community License, which excludes the US, EU, UK and South Korea. Whether you can legally run any of this depends on where you are, and that includes the outputs, not just the weights.