rzgar/minimax_h3_fl2v_lightx2v_4step_int8-convrot_comfy

MiniMax H3 FL2V LightX2V 4-Step Int8-ConvRot

Clip 1: v1 4step | Clip 2: v1 8step <video controls width="1280" height="720"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vlightx2v4stepint8-convrotcomfy/resolve/main/demo/demovl4_8step.mp4" type="video/mp4"> Your browser does not support the video tag. </video>

v0.1 4step demo <video controls width="1280" height="720"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vlightx2v4stepint8-convrotcomfy/resolve/main/demo/MiniMaxH300108_.mp4" type="video/mp4"> Your browser does not support the video tag. </video>

<Gallery />

Model description

Requirements

• —ComfyUI (recent - tested on master 0.31.0)
• —[ComfyUI-LoraInt8Loader](https://huggingface.co/rzgar/minimax_h3_fl2v_lightx2v_4step_int8-convrot_comfy/resolve/main/ComfyUI-LoraInt8Loader/ComfyUI-LoraInt8Loader.zip) custom node (standard ComfyUI LoRA loaders cannot dequantize it)

Files

• —Both are quantized from the KJ comfy-converted BF16 LoRA.

• —LightX2V Release

norefiner variant

The full LoRA also patches the model's tokenrefiner - the text/audio conditioning refiner. The norefiner variant removes those 8 patches so the conditioning path runs at base quality (some video clips had bad audio. This solution fixed it, but keep in mind it doesn’t happen always_). If audio/text conditioning feels weak with the full LoRA at 4 steps, try the norefiner variant.

First clip: with refiner | last clip: norefiner <video controls width="1280" height="720"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vlightx2v4stepint8-convrotcomfy/resolve/main/demo/demo.mp4" type="video/mp4"> Your browser does not support the video tag. </video>

Usage (ComfyUI)

1. 1.Unzip the ComfyUI-LoraInt8Loader custom node into custom_nodes, restart
2. 2.In your MiniMax H3 workflow, add LoRA Loader (Int8-ConvRot), select the file, strength 1.0 (the training scale is already baked in)
3. 3.Sample at 4-6 steps

Download model

Download them in the Files & versions tab.