MiniMax H3 SPEED Calibration Window / 画幅安全标定窗口 (Advanced)

Before you can calibrate anything - a spectrum profile, a SPEED transition, a quality baseline - you need a clean, canonical window of frames: exactly the right length for H3's grid, at the right frame rate, on a canvas that hasn't been mangled. MiniMaxH3SPEEDCalibrationWindowT8Advanced is the tool that cuts that window. It resamples an arbitrary image batch into one strict H3 calibration window: 24fps, the 17n+5 frame count, and an aspect-preserving center-cover resize. It never stretches source geometry, never pads short clips, and - as the description is careful to add - it does not touch your existing Source Media Window. It's the "make me a fair sample" node.

How it works

You feed it frames (an IMAGE batch at whatever fps your source actually is) and tell it the ground truth with source_fps (default 24). Then you declare the target:

• width / height - the canvas (defaults 736×416, stepped in 32s as H3 wants).
• length - the frame count, stepped in 17 from 5 up to 3600, default 124. That's the H3 grid talking: 124 = 7×17+5, a native trained-range length. Type 123 and the stepper snaps you back to a legal value.
• start_seconds - where to begin the cut (0 by default).

The resize_mode is a single option, center_cover: scale the source so it covers the target canvas, then crop the center. That's the aspect-preserving behaviour the description promises - no squashing wide footage into a 3:4 frame, no black bars from padding. If your source is shorter than the window or a different aspect, it still comes out a legal H3 window, but the report tells you exactly what it did.

Outputs: frames (the calibrated window), frame_count (so downstream can confirm), and report_json documenting source fps, target grid, and the resize that happened.

Why this node exists

The whole SPEED/spectrum toolchain is obsessed with comparable inputs. A spectrum profile fit from misaligned windows - wrong fps, off-grid lengths, stretched pixels - is garbage in, and the pack's later nodes are strict enough to notice. This node is the "garbage prevention" stage: one deterministic way to make every clip you feed the Harvester the same shape of thing. If you're using the SPEED feature set at all, this is the front door you should walk through first.

Installing and pairing

Same pack, one install: ComfyUI Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes, restart. No pip deps; the H3 weights are yours. This node is pure CPU image work - no model, no GPU - so it's also the cheapest node in the SPEED family to play with. Typical wiring: frames from a video loader → this node → the Spectrum Harvester (for a probe) or straight into an image conditioning path (for a clean baseline). One practical tip: if your source is 30 or 60fps, set source_fps honestly rather than lying to it - the resample math needs the real number to produce a 24fps window.