Portraits shape first impressions, and lighting is an important factor. Prior work, such as lighting estimation and portrait relighting, edits lighting after capture but does not guide where to place a light before the shot. This leaves an open question: where should a light go to best light a face? Therefore, we propose LightMatch, a preference-grounded inverse lighting framework that disentangles the controllable light from ambient illumination and optimizes its placement. We first learn a Bradley-Terry quality model from pairwise human judgments. A perception head then regresses spherical-harmonic lighting from frozen self-supervised features via cross-attention. Finally, a placement solver isolates the light via spherical-harmonic linearity and optimizes its reachable position. On held-out environments, LightMatch attains 87.8% of the oracle improvement, outperforming the conventional solver that ignores ambient light by a relative 28.4% and the same pipeline with the widely used DPR estimator by a relative 70.5%. It reaches 73.2% on real OLAT faces of unseen identities and, on our real captures, cuts lamp-direction error by 37.7%.
Figures & tables
Figure 2: What Q prefers. (a) Q over lamp position at ρ=2 , with purple to orange representing low to high Q . Each face is a real photograph of a classic setup, placed at its lamp position, and the best of the six, butterfly lighting, is marked by the blue frame. (b) Q from in-house votes and from the votes of four external photographers, at the 30 lamp positions of the Yale faces (dots).
Standard
Real ambient light
Real faces
Real capture
Method
Share ↑
Worse ↓
Recover ↑
Error ↓
Share ↑
Worse ↓
Recover ↑
Error ↓
Share ↑
Worse ↓
Recover ↑
Error ↓
Error ↓
Rating ↑
DECA
4.1
43.9
37.5
68.4
2.6
45.0
10.5
69.4
7.3
41.8
22.5
52.2
101.3
-0.7
Fixed Rembrandt
22.4
34.0
70.0
–
25.1
32.1
26.3
–
25.9
31.4
42.5
–
–
1.1
DPR
51.5
12.5
55.0
42.4
52.9
11.3
31.6
45.0
36.3
24.2
50.0
48.7
50.4
1.8
Ambient-unaware
68.4
9.6
67.5
8.5
71.0
9.4
42.1
7.6
62.4
9.2
57.5
11.5
28.9
2.0
LightMatch
87.8
2.6
75.0
5.7
89.1
2.2
57.9
5.8
73.2
8.6
72.5
8.2
18.0
2.4
Table 1: Lamp placement on the four test sets. For the fourth test set, only the lamp direction is known, so it is evaluated by Error and by Rating.
Figure 3: Lamp-direction error on the fourth test set against the lamp-to-ambient ratio. Each point is one photo. The shaded area marks a lamp twice as strong as the ambient light or stronger. The diamond on each side marks the median error, and the number at the top gives its value.
Figure 4: Share under different movable ranges of the lamp, averaged over the first two test sets. Each fan shows the range of azimuth and elevation within which the lamp can move. Colors at azimuth ±15∘ , ±35∘ and ±75∘ , marked by the white dashed lines and the straight sides of the fan, and at elevation ±10∘ , ±20∘ and ±45∘ , marked by the white solid lines and the outer arc of the fan, are measured, and colors in between are interpolated.
Test set
Variant
Share
Worse
Recover
Error
Standard
LightMatch
87.8
2.6
75.0
5.7
w/o lamp-off
80.4
3.8
67.5
7.3
Real ambient light
LightMatch
89.1
2.2
57.9
5.8
w/o lamp-off
82.3
2.4
52.6
6.6
Real faces
LightMatch
73.2
8.6
72.5
8.2
w/o lamp-off
67.5
9.8
57.5
10.9
Table 2: Ablation on the Standard, Real ambient light and Real faces test sets.
LightMatch
w/ VLM read-out
w/ VLM advice
w/o read-out
GPT-6.1 Sol
Gemini 3.1 Pro
Claude Fable 5.1
Grok 4.7
Qwen3.5-27B
Share ↑
96.2
94.8
94.0
90.5
22.6
25.2
7.5
11.4
0.8
Worse ↓
1.2
1.4
1.4
3.1
31.7
29.7
37.7
32.3
31.5
Table 3: Lamp-placement advice on the no-ambient variant of the first test set. The variant has 800 renders generated in the same way without ambient light. The columns from GPT-6.1 Sol to Qwen3.5-27B are VLMs that give advice directly. Every model writes advice with the same prompt and parser.
Figure 5: Decodability of lighting across the layers of the VLM. Azimuth error of a linear probe on each layer of Qwen3.5-27B, trained and tested on renders. Higher points mean lower error, and layers 31 to 32 read the light best.
Figure 6: Performance and transfer of different vision backbones. (a) Azimuth error. (b) Elevation error. Heads are trained on 2,000 additional renders and tested on 500 held-out renders and on real Yale faces, shown as blue and orange dots, and a longer bar means a weaker transfer from renders to real faces. For both blue and orange dots, further left is better. The dashed line marks the error of guessing the same direction for every image without looking at it, and a point left of the line, that is, with a smaller error, shows that the backbone indeed reads the lighting from the image.
Figure 7: Two examples with real ambient light. (a) An example from the Real ambient light test set with the lamp 6 times the ambient light, shown on a real Yale face. The left map shows Q of the combined light at each lamp position within the movable range, with purple to orange representing low to high Q , gray outside the range covered by human evaluation, and the Share of each method below its face. (b) A real photo from the Real capture test set with the lamp 2 times the ambient light. The leftmost diagram shows the nine lamp positions seen from overhead, and the lamp position is below each photo. The markers show the start, the position chosen by each method, and the best position.
w/o
Ambient-
Fixed
Lamp
LightMatch
lamp-off
unaware
Rembrandt
Error
Standard
Strong
92.1
92.1
69.2
20.0
4.7
Medium
84.9
77.3
72.9
27.2
7.8
Weak
70.0
56.6
52.7
25.0
13.0
Real ambient light
Table 4: Share in percent on the first two test sets, split by the strength of the lamp relative to the ambient light. Error is the lamp-direction error of LightMatch in degrees.
Model
Azimuth
Elevation
Error
Guess
44.5
26.0
52.0
Qwen3.5-27B
43.5
24.4
53.2
Grok 4.7
19.6
20.8
34.0
Gemini 3.1 Pro
17.9
19.6
31.1
Claude Fable 5.1
14.0
20.9
29.9
GPT-6.1 Sol
9.8
14.3
20.1
Table 5: Median lighting error in degrees on the 800 renders of Tab. 3 . Guess is the error of guessing the same direction for every image, as in Fig. 6 .
Fixed
Ambient-
DECA
Rembrandt
DPR
unaware
LightMatch
P1
−0.6
1.2
1.5
1.9
2.3
P2
0.4
1.4
2.1
1.6
2.2
P3
−1.1
0.8
1.9
2.3
2.5
P5
−1.5
1.0
1.7
2.2
2.6
Mean
−0.7
1.1
1.8
2.0
2.4
Table 6: Photographer ratings on the fourth test set. Each value is the mean rating of one photographer on a scale from −3 to +3 , where positive values mean the lighting improves after the move. Photographers P4 and P6 are excluded by the consistency check in Sec. 3 .
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Voting page. Both photos show the same subject, and the photographer clicks the one with better light.
Photographer
Experience
Consistent repeats
P1
3 to 5 years
18/20
P2
5 to 10 years
17/20
P3
more than 10 years
16/20
P4
5 to 10 years
13/20
P5
more than 10 years
17/20
P6
3 to 5 years
9/20
Appendix
Table 7: Experience and repeat consistency of the six photographers.
Figure 9: Q refitted on the votes of each photographer over the tested range for the fill at 1/3 of the frontal image, with purple to orange representing low to high Q and dots at the 30 tested lamp positions. The title gives the experience and the number of repeats with the same choice. The two photographers in gray are discarded.
Condition
Lamp/ambient
LightMatch
Ambient-unaware
Night 2
4.2
6.1
8.1
Night 1
2.8
9.9
10.5
Daylight 4
3.3
15.2
19.6
Daylight 3
7.8
15.3
17.2
Daylight 2
1.0
37.1
63.9
Daylight 1
0.6
42.4
60.5
Appendix
Table 8: Results on the fourth test set by capture condition. Lamp/ambient is estimated from the two photos. The last two columns report the median azimuth error in degrees.
Figure 11: Three more examples from the second test set. The map shows Q of the combined light at each lamp position within the movable range, from purple (low) to orange (high), with gray outside the range covered by human votes. The markers show the start, the position chosen by each method, and the best position. The number below each face is the percentage of the best Q within the movable range, measured from the lowest Q in the range.
Figure 12: Three real photos from the fourth test set. The title of each row gives the lamp height and the lamp strength relative to the ambient light. The diagram on the left of each row shows the nine lamp positions seen from above, with the start and the position chosen by each method, and the lamp position is below each photo.
Both
Azimuth only
Elevation only
Mixed
Test set
Method
±15∘/±10∘
±35∘/±20∘
±75∘/±45∘
±35∘
±75∘
±20∘
±45∘
±35∘/±45∘
±75∘/±20∘
Standard
DECA
50.6
24.8
− 32.8
37.7
− 6.7
28.0
− 14.1
− 12.9
− 13.8
Fixed Rembrandt
44.6
28.1
2.7
24.3
− 3.1
45.4
42.4
27.7
1.3
DPR
62.8
52.2
40.7
54.7
45.6
58.7
53.1
48.8
43.8
Ambient-unaware
84.0
74.1
55.0
77.1
64.8
76.5
64.0
65.0
62.6
LightMatch
92.9
88.1
85.4
89.7
87.1
89.7
85.8
86.4
86.3
Appendix
Table 9: Share (%) under different movable ranges of the lamp. Both enlarges the azimuth and elevation ranges together, written as azimuth/elevation. Azimuth only enlarges the azimuth range with elevation kept at ±10∘ . Elevation only enlarges the elevation range with azimuth kept at ±15∘ . Mixed enlarges one range to the middle setting and the other to the largest setting. All four settings start from ±15∘/±10∘ .
Method
Standard
Real ambient light
Real faces
DECA
5.5
3.7
6.2
Fixed Rembrandt
29.7
28.2
23.4
DPR
60.3
55.1
31.5
Ambient-unaware
66.2
73.6
58.6
LightMatch
83.1
85.9
69.3
Appendix
Table 10: Share (%) computed with the Q refitted on the votes of the four photographers. The lamp positions are those chosen by each method in Tab. 1 .
Method
Standard
Real ambient light
Real faces
Single-lamp optimum
34.3
42.7
33.8
LightMatch
87.8
89.1
73.2
Appendix
Table 11: Share (%) of moving the lamp toward the single-lamp optimum of Q .
Azimuth
Pairs
Percentage (%)
−60∘
15
7.9
−45∘
10
5.3
−30∘
7
3.7
−15∘
11
5.8
0∘
16
8.5
+15∘
15
7.9
Appendix
Table 12: Azimuth recommended by LightMatch on the fourth test set. Each value gives the number and percentage of the 189 pairs for which LightMatch recommends that azimuth.