CompositionEvidence: Controlling Geometric Evidence for Cinematic Composition Reasoning with Frozen Vision-Language Models
Authors: Yichen He, Xin Huang, Yuxi Liu
Organizations: School of Journalism and Communication, Bohai University, Jinzhou, Liaoning, China · School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen), Shenzhen, Guangdong, China
Cinematic composition questions depend on in-frame geometry that vision-language models (VLMs) use inconsistently. External perception tools can expose this geometry, but naive prompt injection can overwrite useful visual predictions. On ShotBench composition, pre-question geometry reduces Qwen2.5-VL-7B from 49.5% to 46.3%; corrupted measurements lower it to 33.6%. CompositionEvidence derives nine reproducible features from frozen detection, depth, and line experts and supplies their shared 24-dimensional representation through two interfaces. A training-free channel appends gated neutral evidence after the question, reaching 56.2% (+6.7, exact McNemar p=0.0018) and 48.4% under corrupted evidence. A minimal-training channel maps the same representation to four soft tokens with a 7.37M-parameter projector while leaving the VLM frozen; it reaches 61.0% (+11.5). Across nine VLMs, including five commercial models queried through official provider APIs, zero-shot accuracy ranges from 43.8% to 63.0%, and the text channel changes it by +6.7 to -6.9 points; the highest baseline is among the regressions. The effect of explicit geometry depends on measurement reliability and on how the receiving model integrates it.
Figures & tables
Figure 1: CompositionEvidence architecture. Frozen detection, depth, and line experts feed deterministic synthesis and nine geometric functions, which yield the shared vector z . Channel A gates and appends selected measurements as post-question text; Channel B maps z into four evidence tokens. Only the fitted thresholds and the projector reach inference.
ID
Quantity
Channel A exposes it when
F1
Subject centroid (x,y)
a dominant subject is detected
F2
Distance to midpoint and thirds
F1 is exposed
F3
Subject area ratio
F1 is exposed
F4
Mirror similarity (SSIM)
horizontal SSIM outside [0.144,0.262]
F5
Left/right visual mass b∈[−1,1]
∣b∣≥0.143
F6
Relative depth layers
a depth map is available
Table 1: The nine geometric features and the Channel-A gates. Channel B always receives the corresponding continuous values and the presence flags.
Condition
Acc. (%)
Δ vs. C0
Exact McNemar
Rule classifier (no VLM)
37.8
−11.7
—
C0 zero-shot (official prompt)
49.5
—
—
Chain-of-thought, no evidence
47.4
−2.1
—
Naive evidence (pre-question)
46.3
−3.1
p=0.276 (75 ↑ /90 ↓ )
Post-question, ungated
53.4
+3.9
—
Channel A: post-question + gates
56.2
+6.7
p=0.0018 (66 ↑ /34 ↓ )
Table 2: ShotBench composition results with Qwen2.5-VL-7B ( n=479 ). Both calibrated channels improve the same zero-shot model; reasoning or raw evidence alone does not. ShotBench reference scores: GPT-4o 55.2, ShotVL-3B 57.4, fine-tuned ShotVL-7B v1.1 62.2.
Figure 2: Confusion matrices for zero-shot (left), Channel A (middle), and Channel B (right). Calibrated evidence recovers directional classes while preserving much of the baseline diagonal; Short side remains nearly absent.
Evidence interface
Control
Acc. (%)
Pre-question, full
none
46.3
Pre-question, numeric
none
45.1
Pre-question, one group
none
48.2
Post-question, full
none
53.4
Post-question, full
strict gates
54.5
Post-question, full
Youden gates
56.2
Table 3: Interface ablation with the evidence source and Qwen2.5-VL-7B fixed. Wording alone does not repair pre-question injection; position supplies the largest step, and selective coverage adds the remaining gain.
Figure 3: A Before/After Q&A demo ( shotbench-2638 ), drawn from the pool of items that C0 answers incorrectly and Channel A answers correctly. The supplementary material adds two further demos and a counterexample in which Channel A breaks a correct answer.
Evidence condition
Acc. (%)
Δ vs. C0
C0: no evidence
49.5
—
Naive: clean
46.3
−3.1
Naive: all fields corrupted
33.6
−15.9
Channel A: clean
56.2
+6.7
position corrupted
54.1
+4.6
visual mass corrupted
51.1
+1.6
Table 4: Clean utility and corrupted-evidence fallback are distinct. Naive injection fails both tests; calibrated rendering gains on clean evidence and returns close to the image-only prior when every field is wrong.
Figure 4: The training-free recipe applied unchanged to nine VLMs on ShotBench composition. The five commercial models are queried through official provider APIs.
School of Arts and Sciences New York University New York City, United States · Dept. of Engineering and Computer Science Oakland University Rochester, United States