Research map / Video / Video generation
Photography Perspective Composition: Towards Aesthetic Perspective Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes photography perspective composition (PPC), generating I2V videos that move from a less favorable to an aesthetically enhanced viewpoint, trained on a 3D-reconstruction-based dataset from expert photos and filtered/aligned by a two-stage VLM quality assessor (PQA) with Flow-DPO.
Problem
Photo composition methods are limited to 2D cropping and fail when subjects are poorly arranged in the scene; there is no perspective-transformation dataset and no defined criteria for perspective quality.
Approach
Treat composition as 3D recomposition by generating camera-perspective transition videos (reversed from expert photos via 3D reconstruction) that guide users toward better views, with a VLM-based PQA model scoring visual quality, motion quality, and composition aesthetics to filter data and drive Flow-DPO alignment.
Key results
- Wan2.1 14B achieves the highest CMM (0.5989) and lowest LPIPS (0.7808) among the three I2V models, with HunyuanVideo I2V best on FVD (264) and PSNR (9.40).
- The PQA model (Qwen2-VL-2B, BTT two-stage) reaches about 0.80-0.81 accuracy on VQ, MQ, and CA with 10 pairs.
- RLHF with Flow-DPO raises CMM from 0.4928 to 0.5014 and CA from 0.7070 to 0.7342 on the HunyuanVideo-based model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add RLHF (ablation) | PPC with Hunyuan I2V · PPC videos | CMM: 0.4928 -> 0.5014 | Helped |
| add RLHF (ablation) | PPC with Hunyuan I2V · PPC videos | FVD: 264.7672 -> 270.2212 | Hurt |
| add RLHF (ablation) | PPC with Hunyuan I2V · PPC videos | VQ: 0.7216 -> 0.7477 | Helped |
| add RLHF (ablation) | PPC with Hunyuan I2V · PPC videos | MQ: 0.7496 -> 0.7774 | Helped |
| add RLHF (ablation) | PPC with Hunyuan I2V · PPC videos | CA: 0.7070 -> 0.7342 | Helped |
| increase training data from 20% to 40% (ablation) | PPC with Wan2.1 14B · PPC videos | CMM: 0.5014 -> 0.5989 | Helped |
| increase training data from 20% to 40% (ablation) | PPC with Wan2.1 14B · PPC videos | FVD: 460 -> 345 | Helped |
| increase training data from 40% to 80% (ablation) | PPC with Wan2.1 14B · PPC videos | CMM: 0.5989 -> 0.5244 | Hurt |
| increase training data from 40% to 80% (ablation) | PPC with Wan2.1 14B · PPC videos | FVD: 345 -> 362 | Hurt |
| train with 20° instead of 10° rotation angles (ablation) | PPC with Wan2.1 14B · PPC videos | CMM: 0.4413 -> 0.5587 | Helped |
| train with 20° instead of 10° rotation angles (ablation) | PPC with Wan2.1 14B · PPC videos | FVD: 397 -> 337 | Helped |
| train with 30° instead of 20° rotation angles (ablation) | PPC with Wan2.1 14B · PPC videos | CMM: 0.5587 -> 0.3983 | Hurt |
Surprising findings
- Perspective transformations of 30 degrees degrade both accuracy and quality versus 10 and 20 degree data.
- Model outputs become unstable as training data volume grows beyond a point, with performance deteriorating.
Limitations
- Video duration is constrained by existing video generation models.
- Training data generated by 3D reconstruction models limits quality of the perspective transitions.
- Model outputs become unstable as training data scales.
Experimental setup
Datasets: GAICD SACD FLMS FCDB Unsplash
Benchmarks: VBench 2.0 I2V metrics FVD LPIPS PSNR SSIM
Metrics: CMM FVD PSNR SSIM LPIPS CLIP-F VQ MQ CA accuracy
Baselines: CogVideoX 1.5 5B HunyuanVideo I2V Wan2.1 14B
Builds on: ViewCrafter Diffusion-DPO VideoAlign Bradley-Terry with ties (BTT) VBench
Closest papers by meaning
- The Photographer's Eye: Teaching Multimodal Large Language Models to See, and Critique Like Photographers (CVPR 2025 · similarity 0.84)
- Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping (CVPR 2026 · similarity 0.84)
- Can Machines Understand Composition? Dataset and Benchmark for Photographic Image Composition Embedding and Understanding (CVPR 2025 · similarity 0.82)
- MMPerspective: Do MLLMs Understand Perspective? A Comprehensive Benchmark for Perspective Perception, Reasoning, and Robustness (NeurIPS 2025 · similarity 0.82)
- Generative Photography: Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis (CVPR 2025 · similarity 0.82)
- GenRec: Knowing Where to Reconstruct and Where to Generate (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.