"This green won't key" was wrong — measure green strength, not hue

NaoVerde · 2026-09-03

Conditions tried

5

Green at 3:4

hue 122 / L 0.91

Green at 9:16

hue 150 / L 0.32

Edge after keying

1.26%

Correction — the conclusion of this article was wrong (2026-09-05)

This article originally concluded "in vertical we stopped keying by colour." That was wrong. The green produced at 9:16 (hue 150 degrees, lightness 0.32) keys cleanly once you change how you key it. Thinking in terms of After Effects' Keylight, we measured hue and lightness to decide whether the colour worked as a key. What actually matters is spill = G - max(R,B): how far green exceeds the other channels at each pixel. A green can be 30 degrees off and dark and still key, as long as it dominates. Applying that to the same 14 frames from condition D leaves 1.26% semi-transparent edge on average, with no fringe visible. Everything else in this article still holds — stretching to vertical makes the model paint a scene, the reference decides composition and the prompt decides background. Only the final conclusion changed.

Changing the aspect ratio changes how the model treats the background

Same prompt, same reference, same character; only the resolution changed. At 3:4 (608x800) the green holds to the last frame. At 9:16 (608x1088) the model throws the green away and paints a cityscape. The dance and the character are intact — only the background becomes something else. The model fills the extra vertical space with scenery. Training data at 9:16 skews toward vertical short-form video, where a scene is the norm.

Four conditions, laid out horizontally

Vertical video is easiest to compare when the frames run left to right in one row. A 4x4 contact sheet folds time back on itself and hides the transition. Each row is one video, 14 frames sampled evenly from 124. Condition A (flat green reference plus green instruction) went black from the second frame. Condition B (flat green reference plus a scene instruction) produced the scene but left 44 of 124 frames green. Condition C (reference composited onto the scene) reduced the leak to 15 frames — and the remainder was our own mistake: the template substitution had only applied in one place, leaving four mentions of green in the prompt body. Condition D, run with those mentions still in place, produced green in all 124 frames.

What we got wrong, and what to measure instead

At that point we reported that vertical chroma key was solved, and were told the green was far too dirty for it. We measured: hue 150 degrees against a target of 120, lightness 0.32. We read that as unusable and stopped keying by colour. The measurement was fine; the choice of what to measure was not. Hue and lightness describe how a human sees the colour. A key needs to separate foreground from background, and for that what matters is dominance: spill = G - max(R,B). Compute alpha as clip((60 - spill) * 6, 0, 255), then for pixels where spill > 0 pull G down to max(R,B) to remove the green cast on the edges. Four lines. It works on dark, off-hue green.

Conclusion (corrected) — vertical keys fine; we measured the wrong thing

Green keys at 9:16. We did not need to give it up. There were two separate obstacles: the first (the model would not produce green at all) was a real problem, solved by making the reference and the prompt agree. The second (the green we did get looked dirty) was our measurement, not the footage.

Revision history

2026-09-05 — Conclusion corrected. "In vertical we stopped keying by colour" was wrong: the green produced at 9:16 (hue 150, lightness 0.32) keys cleanly with spill = G - max(R,B). Measuring hue and lightness was the mistake. Evidence figure added and the title changed. Everything else in the article still holds.

A note on the MiniMax H3 licence

Why this page carries no MiniMax H3 output (2026-09-15). The MiniMax H3 Community License limits the licence to an Applicable Territory — "worldwide, excluding the Excluded Territories" — and names those as the European Union, the United Kingdom, the Republic of Korea and the United States of America (§I.3, §I.5). §III.4 forbids using, reproducing, distributing or displaying the Works or any of their Outputs outside that territory. This English edition is read from those places, so we removed the H3 clips and stills from it. The measurements stay — they are ours, not outputs of the model. Licence text: https://huggingface.co/MiniMaxAI/MiniMax-H3/raw/main/LICENSE

Related

• MiniMax H3 ref2va with the 4-step Turbo LoRA: 16.1 min to 3.8 min at 1344x768
• Is FLUX really non-commercial? We counted 312 car and train LoRAs

MiniMax H3 chroma key vertical video prompt design ComfyPods