Recent generative image-editing Diffusion Transformers (DiTs) demonstrate impressive semantic editing capabilities but still struggle with spatially consistent camera angle changes. A primary bottleneck in training foundation models to execute free-form, promptable camera angle changes is the lack of specialized training data. While multi-view datasets exist for generic 3D environments and objects, there remains an absence of paired, multi-view datasets featuring human subjects at fixed locations in natural scenes, including frontal and side-profile views. Capturing such multi-camera data in unconstrained environments is logistically challenging and unscalable. In this paper, we first experiment with multiple state-of-the-art image editing models to create this data synthetically, but find that the outputs are frequently prone to hallucinations involving how much the subject's head turns relative to the background, often producing inconsistent environments. To address this issue, we propose the Head Scene Rotation Difference (HSRD) metric to quantitatively evaluate camera movements around a person. The proposed metric operates by decoupling camera movement from localized head pose manipulation. As demonstrated by the extensive experimentation, HSRD provides the pipeline necessary to evaluate 3D spatial parallax for a person in a scene, paving the way to reliably construct high-quality multi-view synthetic datasets.
Figures & tables
Figure 2: Sample of frontal human scene dataset. A representative sample of the synthetic 40 image dataset generated from Qwen-Image-2512 including the following examples: (top left) Woman in Bedroom, (top right) Man in Kitchen, (bottom left) Woman in Kitchen, and (bottom right) Man in Laundry Room.
Configuration
PSNR ↓
MSE ↑
LPIPS ↑
GPT Image 2 (P1)
22.8908
0.013336
0.4538
GPT Image 2 (P2)
22.2274
0.016670
0.4416
GPT Image 2 (P3)
13.6919
0.055676
0.5761
Nano Banana 2 (P1)
27.4104
0.011071
0.4162
Nano Banana 2 (P2)
24.0676
0.024641
0.4469
Nano Banana 2 (P3)
22.4992
0.026416
0.4444
Table 1: Quantitative evaluation of background shift for detecting camera angle change in generated images. Performance is aggregated across various models and prompt variations, as well as the Qwen camera angle change LoRA. Metrics are computed over a fixed evaluation subset of 80 generated images per group. Metric arrows indicate direction of lower background similarity.
Model
Mode
AUC@30 ( ↑ )
Yaw Error ( ↓ )
Pitch Error ( ↓ )
Roll Error ( ↓ )
VGGT
Original (Unmasked)
88.33
1.21 ∘
0.52 ∘
0.98 ∘
VGGT
Foreground Masking
86.93
1.54 ∘
0.66 ∘
1.11 ∘
MASt3R
Original (Unmasked)
77.63
7.05 ∘
2.31 ∘
2.15 ∘
MASt3R
Foreground Masking
76.15
8.29 ∘
2.37 ∘
2.22 ∘
DUSt3R
Original (Unmasked)
71.19
9.76 ∘
2.98 ∘
2.14 ∘
DUSt3R
Foreground Masking
69.15
10.63 ∘
3.26 ∘
2.16 ∘
Table 2: Quantitative impact of foreground segmentation on camera pose estimation. Evaluation of background tracking stability on the RealEstate10K dataset. Errors are reported in degrees. Metric arrows denote direction of better performance.
Subset
Mean / Std
Q1
Median
Q3
All Views
47.05∘±44.83∘
8.70∘
30.52∘
73.35∘
Left Views
34.72∘±35.35∘
6.35∘
16.61∘
61.38∘
Right Views
59.37∘±49.66∘
14.20∘
63.64∘
84.56∘
Table 3: HSRD distribution and quartile analysis by view direction. Breakdown of HSRD metrics across the overall dataset and directional subsets, highlighting the skew toward higher rotational discrepancy in right-ward prompts.
Figure 3: HSRD analysis of camera LoRA output images. Input images on the left for each pair are labeled with the requested camera change direction, such that the head in the output image should face opposite to the labeled direction. LoRA output images are labeled with their corresponding HSRD scores. Images with lower HSRD scores show logically consistent camera angle changes, while images with higher HSRD scores near Q3 tend to exhibit a frozen background with only the head turning to simulate the requested camera change.