Research map / Image generation / Text-to-image diffusion
PreciseCam: Precise Camera Control for Text-to-Image Generation
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 PreciseCam, a ControlNet-based adapter for SDXL that conditions text-to-image generation on four camera parameters (roll, pitch, vFoV, distortion ξ) encoded as Perspective Field–Unified Spherical maps, trained on a new 57,380-image dataset built from 360° panoramas.
Problem
Current text-to-image diffusion models lack precise control over camera viewpoint and lens distortion, and prompt engineering or predefined shot tags give limited, unreliable control.
Approach
Encode user-specified roll, pitch, vertical FoV and distortion ξ as per-pixel Perspective Field maps (up-vectors and latitudes under the Unified Spherical camera model) and inject them into SDXL via a ControlNet, using only single-view images with no 3D geometry or multi-view data; at inference only the middle-block residuals are injected.
Key results
- Prompt adherence comparable to SDXL with prompt engineering: CLIP 0.543 vs 0.557 and BLIP 0.292 vs 0.299 (per Fig. 7 table).
- User study with 34 participants found the model significantly outperforms SDXL and Adobe Firefly in matching the desired camera view.
- Ablation shows injecting ControlNet residuals only at the middle block keeps camera adherence while preserving SDXL image quality; deeper injections introduce distortions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Condition SDXL generation on PF-US camera maps instead of using prompt engineering alone | SDXL text-to-image generation · 2,940 generated images | CLIP score: 0.292 (0.026) -> 0.299 (0.025) | No clear effect |
| Condition SDXL generation on PF-US camera maps instead of using prompt engineering alone | SDXL text-to-image generation · 2,940 generated images | BLIP score: 0.557 (0.039) -> 0.543 (0.039) | No clear effect |
| Inject ControlNet residuals only at the bottleneck (ablation) | PreciseCam · Camera-conditioned image generation | camera conditioning adherence | Helped |
| Inject ControlNet residuals only at the bottleneck (ablation) | PreciseCam · Camera-conditioned image generation | image quality | Helped |
| Also inject ControlNet residuals at additional levels (ablation) | PreciseCam · Camera-conditioned image generation | image quality | Hurt |
Surprising findings
- Injecting ControlNet residuals only at the bottleneck at inference, rather than at all levels used in training, improves consistency without hurting condition adherence.
Limitations
- Struggles with extreme roll because the backbone tends to align objects vertically.
- Conflicts between prompt and desired camera view can produce semantically incoherent outputs.
- Dataset prompts are generated by BLIP-2, which may be inaccurate.
Experimental setup
Datasets: 360-SOD CVRG-Pano F-360iSOD Poly Haven HDRIs Sitzmann et al. 360 dataset 360cities
Benchmarks: CLIP score and BLIP score on 2,940 generated images user study with 34 participants
Metrics: CLIP score BLIP score user preference for camera view alignment perspective-field-based camera view error
Baselines: SDXL with prompt engineering Adobe Firefly Image 3 Kumari et al. custom diffusion with 3D object conditioning Voynov et al. universal lens geometry
Builds on: ControlNet Perspective Fields Unified Spherical camera model Stable Diffusion XL
Closest papers by meaning
- Generative Photography: Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis (CVPR 2025 · similarity 0.90)
- Large-Scale Text-to-Image Model with Inpainting is a Zero-Shot Subject-Driven Image Generator (CVPR 2025 · similarity 0.87)
- Zero-Painter: Training-Free Layout Control for Text-to-Image Synthesis (CVPR 2024 · similarity 0.87)
- FineStyle: Fine-grained Controllable Style Personalization for Text-to-image Models (NeurIPS 2024 · similarity 0.87)
- Focus-N-Fix: Region-Aware Fine-Tuning for Text-to-Image Generation (CVPR 2025 · similarity 0.87)
- XVerse: Consistent Multi-Subject Control of Identity and Semantic Attributes via DiT Modulation (NeurIPS 2025 · similarity 0.86)
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.