Free ComfyUI workflow for MiniMax H3 text-to-video, made alongside the Frozen World LoRA (trigger timefr33ze).
What it does
Modes: 1 preview (1024x576, 73 f, 20 steps) · 3 final (1344x768, 107 f). Mode 2 (fast / Turbo) is UNTESTED with Frozen World.
Scene helper: type the frozen things in SCENE; pick a MOVE (1-4 camera moves, 5-6 walker) and a TRIGGER STYLE (3 = no trigger, for base comparison).
LoRA slot 1: FrozenWorld_v1.safetensors @ 1.0. Bypass it / set strength 0 to run base H3.
Guidance locked at 1 (CFG-distilled). Audio is generated with the video.
Frozen World tip
Keep the camera STILL (camera moves weaken the freeze). Prompt: "timefr33ze, photorealistic cinematic footage, the entire world is frozen in time." + list the frozen things + "Only <one person> moves, walking toward the camera; the camera holds perfectly still." Tested at 1024x576, 107 frames.
Hardware
Tested on 16 GB (RTX 5080) with int8_convrot weights and on 24 GB. 12 GB GGUF branch is untested.
Install
Load the .json in ComfyUI and install missing custom nodes with ComfyUI-Manager.
Put FrozenWorld_v1.safetensors in ComfyUI/models/loras.
Set MODE, MOVE, TRIGGER STYLE and SCENE, then Queue.