kinetic_cuts

A MiniMax H3 video LoRA (rank 16) trained with ai-toolkit on 12 video clips of up to 1.62 s (39 frames at 24 fps) with their audio, cut evenly from 1 source video(s) with a 10% edge crop for 380 steps in 41 minutes on one A100. It was trained to pick up the look, the motion and editing over time, and the sound (captions named only subjects, actions and on-screen words, so everything else binds to the trigger).

Reference clip: “A clown at the circus rides on a zebra, kinetic_cuts style” (10 s, seed 4242, same prompt for every LoRA in this series).

Trigger

Start the prompt with kinetic_cuts — the captions were kinetic_cuts, …, and H3 LoRAs follow the caption words, not the bare trigger token.

Use in ComfyUI

Load kinetic_cuts.safetensors with LoraLoaderModelOnly after the H3 diffusion model (minimax_h3_fl2va_pruned_bf16) and its turbo LoRA; strength 1.0. Works with text-to-video, first/last-frame and reference-to-video graphs. The keys are the ComfyUI layout (diffusion_model.blocks.N…lora_A/lora_B), the same as fal's published H3 LoRAs.

Samples (final step, 39-frame clips)