Both fixtures read as live sources. Warm pools land on the tables, the wall and both faces; the room edges fall away. This is the "practicals, warm" look you asked for on the H3 lane, and it is the only arm where the light source and the light agree.
Lamps stay dark; the light comes from an unseen cool source. Least warm of the set, most obviously night by colour. Good for a cold, lonely read.
Strong orange practical on the right — but it also paints blue-violet streaks across the upper wall that exist nowhere in the plate. A new lighting effect rather than a relight. Brightest of the set at 0.76× the day plate.
No look steer at all: the model's own idea of night. Darkest at 0.50×, one lamp glowing, highlights well contained. The floor the looks are built on.
| arm | mean luma | p50 | p99 | p99.9 | vs day plate | cast shift |
|---|---|---|---|---|---|---|
| day plate | 63.1 | 57.9 | 186.5 | 208.1 | — | — |
| base · neutral | 31.5 | 27.1 | 116.4 | 151.7 | 0.50× | −37.0 |
| L1 · both lamps | 42.1 | 28.4 | 153.2 | 205.7 | 0.67× | −11.8 |
| L2 · warm tungsten | 47.8 | 32.9 | 169.5 | 206.6 | 0.76× | −34.5 |
| L4 · cool moonlight | 35.9 | 28.3 | 127.1 | 159.4 | 0.57× | −77.1 |
Cast shift is how far the red-blue balance moved from the day plate. L1 barely moves (−11.8) — the warmth that was in the plate is still there. L4 moves furthest, which is exactly what a moonlight look should do.
Same L1 look, 5 seconds and 8 seconds, one continuous pass — no chunks, no crossfade, no seams. Motion is scored against the source plate's own frame-to-frame movement at that exact moment; a perfect relight scores 1.00× with correlation 1.00.
| length | motion ratio | correlation | luma across the clip | verdict |
|---|---|---|---|---|
| 97f · 4.0s | — | — | — | reference |
| 121f · 5.0s | 0.85× | +0.736 | 39.5 → 38.1 → 37.2 | passes |
| 193f · 8.0s | 0.79× | +0.538 | 40.8 → 41.4 → 42.6 | passes |
The camera move survives both. At 193f the light settles slightly in the first second (39 → 33) and then holds flat at 33–36 for the remaining seven — that is the camera moving through the room, not a fade. There is no banding: measured row-to-row discontinuity is 2.9× the median, at the bottom edge of frame.
No arm invented a phone light, a match, or a new lamp. The only alteration is ordinary detail drift — hands, towel folds — not new illumination. That was the failure mode on the H3 lane.
This is a full-length window, not a 39-frame clip below the model's minimum. The performance does not get filled in with invented action.
The failure to watch is an arm that names bright things too enthusiastically — L2's wall streaks are what that looks like. One clean sentence per look beats a block of retention text.
~4½ min and a few cents per arm at 768×448. Four looks in one sitting, which is why the comparison is worth doing properly.