Qifeng Chen

Capturing compelling photographs requires sequential capture-time decisions. However, existing guidance methods operate at the image level rather than modeling the decision process that diagnoses photographic dimensions (e.g., camera position, lighting) and resolves them into executable reshooting actions, offering limited actionable support for novices. We present PhotoMaster, a decision-driven multimodal framework for capture-time photography guidance. To model the decision-making process, we propose the Photography Decision Taxonomy (PDT) that defines photographic dimensions for diagnosing novice-captured photos and maps them to executable actions. To make this intuitive, PhotoMaster pairs these executable actions with a rendered visual target that previews the expected outcome. To learn this decision space, we introduce a two-fold data engine: Forward Decision Transfer distills professional photographers' decisions from masterworks to casual shots, while Decision Inversion applies inverse capture decisions via camera-controllable world models, collectively yielding over 230K PDT-annotated triplets. Evaluations demonstrate PhotoMaster's capability to produce expert-aligned diagnoses and highly preferred visual references, while a real-world user study confirms that novices achieve significantly better captures using our framework.

We present LiveLight, the first diffusion-based framework for real-time streaming video relighting with interactive 3D lighting control. Achieving this is non-trivial, as it requires overcoming three critical challenges: effectively injecting dynamic 3D lighting into a diffusion model, maintaining high-fidelity generation under an extremely low NFE (Number of Function Evaluations) budget for real-time speed, and facilitating continuous streaming for interactive control. To address these pain points, we propose three key designs. First, for accurate lighting injection, we propose a lightweight adapter that feeds Multi-Plane Light Irradiance (MPLI) conditions—depth-aware irradiance maps encoding 3D lighting geometry—directly into the diffusion backbone. Second, to prevent rendering quality degradation at low NFEs towards real-time distillation, we introduce a geometry-guided feedback branch. This training-time constraint leverages a frozen geometry estimator to enforce depth- and normal-consistent relighting, ensuring geometrically plausible shading without adding inference overhead. Finally, to enable streaming interaction, we develop a progressive rolling-window strategy that maintains a denoising ladder of latent chunks at varying noise levels. By propagating intermediate states, this strategy guarantees temporal coherence and supports arbitrarily long video relighting with per-frame reference refresh. Extensive experiments on real-world and synthetic benchmarks demonstrate that LiveLight achieves state-of-the-art relighting quality while running at real-time speed, significantly outperforming offline baselines in temporal stability, lighting controllability, and user preference. To foster real-time interactive relighting research, we will publicly release our models, training data, and synthetic data generator on www.dummy.url.