Generative Photography: Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis

TL;DR

Generative Photography introduces a space-camera joint paradigm to enable scene-consistent text-to-image synthesis under varying camera intrinsics. The framework combines Dimensionality Lifting, which shifts from pure spatial generation to a space-camera representation via a T2V backbone, with Differential Camera Intrinsics Learning, which uses differential data and a dedicated encoder to model frame-to-frame camera setting changes. Through carefully constructed differential data, physically grounded rendering, and a CLIP-informed encoder, the approach achieves higher accuracy in aligning with camera physics and maintains stronger scene consistency than state-of-the-art baselines. The results indicate practical potential for camera-aware generative tools in professional photography, reducing post-processing and enabling photorealistic, controllable image synthesis.

Abstract

Image generation today can produce somewhat realistic images from text prompts. However, if one asks the generator to synthesize a specific camera setting such as creating different fields of view using a 24mm lens versus a 70mm lens, the generator will not be able to interpret and generate scene-consistent images. This limitation not only hinders the adoption of generative tools in professional photography but also highlights the broader challenge of aligning data-driven models with real-world physical settings. In this paper, we introduce Generative Photography, a framework that allows controlling camera intrinsic settings during content generation. The core innovation of this work are the concepts of Dimensionality Lifting and Differential Camera Intrinsics Learning, enabling smooth and consistent transitions across different camera settings. Experimental results show that our method produces significantly more scene-consistent photorealistic images than state-of-the-art models such as Stable Diffusion 3 and FLUX. Our code and additional results are available at https://generative-photography.github.io/project.

Figures (12)

• Figure 1: (a) Existing text-to-image (T2I) models struggle to perceive physical camera settings and maintain consistency across multiple settings, even when the random seed is fixed. (b) We solve this problem by lifting camera-controlled text-to-image (T2I) generation into text-to-video (T2V) generation, thereby decoupling scene description from camera settings and achieving better scene consistency.
• Figure 2: The pipeline of building differential data.
• Figure 3: The overall architecture of differential camera encoder.
• Figure 4: Visual comparisons between different generative methods. Our method is capable of generating realistic camera effects for any given camera setting scale, while maintaining high scene consistency across images corresponding to different scales. Both AnimateDiff Guo_2023_AnimateDiff and CameraCtrl He_2024_Cameractrl have been fine-tuned on our data. We highlight the discontinuities in the scene with red arrows.
• Figure 5: Ablation study on dataset scaling characteristic.