CamEdit: Continuous Camera Parameter Control for Photorealistic Image Editing

We sincerely appreciate your positive feedback and comments. Below are our responses to your concerns:

W1&Q1: Concern about whether editing three camera parameters requires training three separate models.

We conduct an experiment training a unified model to adjust all three parameters simultaneously. In this setting, we doubled the number of training parameters and incorporated 5,000 synthetic images with multi‑parameter variations. As shown in Table R4‑1, the unified model achieves accuracy and output quality comparable to single‑parameter control, confirming its effectiveness in handling multi‑parameter adjustments. We will include these updated results and further clarify this capability in the revised manuscript.

Table R4-1: Performance of CamEdit with single-parameter vs. joint multi-parameter control, evaluated across 200 images with various parameter settings.

W2&Q2: Request for results on a larger benchmark.

We expand the testing set to 200 images. As shown in Table R4‑2, CamEdit consistently outperforms all baselines across aperture, focal plane, and shutter speed, achieving superior perceptual quality, semantic consistency, and parameter accuracy on all tasks. These results are consistent with those reported in the main paper and will be incorporated into the revised version.

Table R4‑2: Quantitative comparison of 200 images across aperture, focal plane, and shutter speed.  

W3&Q3: More exploration and discussion about the current disability of inverse recovery.

We clarify that our method—similar to most camera parameter rendering approaches (e.g., DrBokeh [1], MPIB [2]) and image editing frameworks (e.g., SuperEdit [3])—is not designed to restore information that is already absent from the input. CamEdit with the proposed modulation simulates the effects of adjusting camera parameters but does not attempt to reconstruct content that was never captured. This limitation arises because our task focuses on photographic editing, where fidelity to the original content is prioritized over new content generation.

However, we recognize the value of exploring inverse image restoration for image editing. Our CamEdit50K pipeline provides a foundation for such research, and we plan to investigate extending our framework for restoring degraded content in future iterations.

[1] Yichen Sheng et al. Dr. Bokeh: Differentiable Occlusion-Aware Bokeh Rendering. CVPR, 2024.

[2] Juewen Peng et al. MPIB: An MPI-based bokeh rendering framework for realistic partial occlusion effects. ECCV, 2022.

[3] Fan Chen et al. SuperEdit: Rectifying and Facilitating Supervision for Instruction-Based Image Editing. ICCV, 2025.

Dear Reviewer U816,

We sincerely appreciate your positive feedback, and have addressed your concerns accordingly. As the second review phase nears its end, please let us know if anything remains unclear or if further clarification would be helpful.

We sincerely appreciate your positive feedback and comments. Below are our responses to your concerns:

W1: About Typos.

We will correct all typos in the revised manuscript.

W2: Clarification on how continuous parameter prompting and parameter-aware modulation connect to the problem.

Existing diffusion-based editing frameworks face two key limitations when applied to photorealistic, parameter-controllable image editing:

1. Camera parameters are inherently continuous, whereas pre‑trained diffusion backbones (e.g., SD3) encode textual prompts as discrete tokens through their text encoders. As a result, most diffusion‑based editing methods struggle to capture smooth parameter adjustments, limiting their applicability to photographic editing tasks.

2. Diffusion models lack prior knowledge of camera parameters and fail to capture the corresponding visual effects for each specific parameter.

To address these limitations, we introduce two complementary modules:

• Continuous Parameter Prompting interpolates between semantically meaningful anchor embeddings within the continuous text embedding space. This preserves alignment with representation distribution of the pre-trained model while enabling stable, fine-grained, and perceptually coherent control over camera parameters, as demonstrated (see Sec. 5.2).
• Parameter-Aware Modulation moulates the feature inside the diffusion transformer to explicitly capture spatial and global effects that text embeddings alone cannot represent. It applies spatial warping to model geometry/focus changes and channel‑wise scaling to account for exposure and intensity variations.

By integrating these two modules, our framework enables stable, realistic, and continuous camera parameter editing. We will refine Sec. 1 to clarify these connections.

W3: Request for more discussion on the dataset, including the pros and cons of data sources, the effect of data ratios, and the influence of scale.

Real paired data provide faithful camera effects and realistic details, helping the model learn real parameter-dependent visual cues. However, collecting large-scale paired real data is challenging, particularly for diverse conditions such as animals or uncontrolled outdoor scenes. Synthetic data, on the other hand, enables broader coverage of parameter values and scene diversity while ensuring precise annotations.

We evaluate the impact of these data sources through an ablation study (Table R3-1).  

• Training only on real data (~7k) results in lower perceptual quality and parameter accuracy due to the limited diversity of real scenes and insufficient parameter coverage.
• Adding synthetic data, with a 1:1 synthetic-to-real ratio (~15k), consistently outperforms training with real data alone. However, the effect is limited by the overall dataset size.
• Training only on synthetic data (~40k) improves the image content and parameter control due to its large quantity. However, its visual quality is limited due to the lack of guidance from real captured images.
• Full CamEdit50K, by combining both available synthetic and real data, achieves better parameter control and visual quality. This improvement is driven by both the quantity of synthetic data and the quality of the real data. As shown in Figure 6 of the main paper, when compared with the rendering-based method (BokehMe) used for our synthetic data, our CamEdit significantly improves visual results, reflecting the effectiveness of combining real and synthetic data.

These findings demonstrate that both the diversity and scale of the dataset are critical for maximizing performance. We will incorporate these discussions and the ablation results into the revised manuscript.

Table R3-1: Performance of CamEdit with different synthetic-to-real data ratios.

W4: Concern about evaluation reliability due to the use of GPT-generated samples.

We clarify that our evaluation is not based on GPT‑generated samples. All quantitative experiments are conducted on real‑captured images. In addition, we include both a human user study and benchmarks on real datasets with ground‑truth captures to further validate the reliability of our results.

1. Benchmarks on Real Datasets (Table R3‑2).   We benchmark CamEdit on three publicly available real datasets containing ground‑truth captures from digital cameras and photographic lenses: VABD [1] for aperture, EBB Val [2] for focal plane, and SICE [3] for shutter speed. As shown in Table R3‑2, CamEdit outperforms both image editing methods fine‑tuned on CamEdit50K (UltraEdit*) and rendering‑based approaches (DrBokeh, CLODE) across SSIM and LPIPS for all tasks.

2. User Study (Table R3‑3). We conduct a user study with 15 participants who have expertise in photography. Each participant evaluates 20 image sets per task across all three tasks, scoring photographic realism, parameter accuracy, and content preservation on a 0–4 scale. As shown in Table R3‑3, CamEdit consistently achieves the highest scores across all metrics, confirming that our results remain valid under human evaluation.

These results demonstrate that our evaluation is grounded in both real-world data and human judgments, ensuring the reliability of our reported performance.

Table R3-2: Quantitative comparison of CamEdit on three real datasets (VABD for Aperture, EBB for Focal Plane, and SICE for Shutter Speed).

Table R3-3: User Study Scores (0–4 scale) on Photographic Realism, Parameter Accuracy, and Content Preservation

[1] Andrey Ignatov, et al. AIM 2020 Challenge on Rendering Realistic Bokeh. ECCVW, 2020.

[2] Kang Chen, et al. Variable Aperture Bokeh Rendering via Customized Focal Plane Guidance. arXiv, 2024.

[3] Jianrui Cai, et al. Learning a Deep Single Image Contrast Enhancer from Multi-Exposure Images. TIP, 2018.

L: Request to discuss the limitations of the work.

As mentioned in Section 6 of the main paper, our work already covers some camera parameters, but we plan to extend the approach to include more parameters and editing tasks in future work. We will explicitly outline these plans and limitations in the revised manuscript.

Dear Reviewer DyLt,

We sincerely appreciate your positive feedback, and have addressed your concerns accordingly. As the second review phase nears its end, please let us know if anything remains unclear or if further clarification would be helpful.

We sincerely thank you for your review. Below are your concerns and our corresponding responses:

W1: Analysis of dataset composition and the accuracy of estimated real-image parameters.

We conduct two sets of ablation studies to address these concerns.

1. Impact of CamEdit50K composition (Table R2-1).
We train models on only real data, only synthetic data, and the full CamEdit50K dataset.

• Training only on real data (~7k) results in lower perceptual quality and parameter accuracy due to the limited diversity of real scenes and insufficient parameter coverage.

• Training only on synthetic data (~40k) achieves more consistent parameter estimation due to its large quantity. However, its visual quality is limited due to the lack of guidance from real captured images.

• Full CamEdit50K, by combining both available synthetic and real data, achieves better parameter control and visual quality. This improvement is driven by both the quantity of synthetic data and the quality of the real data. As shown in Figure 6 of the main paper, when compared with the rendering-based method (BokehMe) used for our synthetic data, our CamEdit significantly improves visual results, reflecting the effectiveness of combining real and synthetic data.

Table R2-1: Performance of CamEdit with different dataset composition.

2. Effect of estimated camera parameters (Table R2-2).
We compare models trained with synthetic data combined with either only EXIF-annotated real data, only estimated-parameter real data, or the full dataset.

• Training on estimated-parameter data improves perceptual quality relative to EXIF-only data, as the estimated subset spans more diverse scenes. Real data with estimated parameters also contribute to improvements in both the visual quality and parameter control.
• The full dataset achieves the best overall performance, confirming the accuracy of our estimation method. Estimated parameters allow the inclusion of more real paired data into the dataset, further enhancing the model’s performance.

These results demonstrate that both real and synthetic data are essential for optimal performance, and our estimated camera parameters improve coverage and help the model generalize, ensuring reliable editing quality.

Table R2-2: Impact of estimated vs. EXIF camera parameters on editing performance in aperture task.

W1: Design choice of varying one parameter at a time vs. multiple parameters.

We conduct an experiment where the model is trained to adjust all three parameters simultaneously. In this experiment, we double the number of training parameters and include 5,000 synthetic images with multi-parameter variations. As shown in Table R2‑3, the unified model demonstrates accuracy and output quality comparable to single-parameter control, confirming its ability to effectively manage multi-parameter adjustments. We will incorporate these updated results and provide further clarification on this capability in the revised manuscript.

Table R2-3: Performance of CamEdit with single-parameter vs. joint multi-parameter control, evaluated across 200 images with various parameter settings.

W2: Clarity of Section 4.2 and the description of the Parameter-Aware Modulation module.

In the revised version, we will:

1. Improve the method figure by explicitly depicting the two components of the parameter-aware modulation module and their roles within the diffusion backbone.

2. Clarify terminology by replacing the shorthand “Down” with the standard term “AvgPool (2×2)” to avoid confusion.

3. Explain grid_sample with a concise description of its operation and purpose. In our framework, it is used to warp features according to the camera parameter offsets, enabling spatially aligned modulation of the diffusion features.

We will enhance the readability of Section 4, ensuring that the parameter-aware modulation module is easy to understand.

W3: Baseline comparison for the Parameter-Aware Modulation (PAM) module.

We conduct more ablation studies to demonstrate the effectiveness of our PAM.

• Replace by TimeStep: injecting continuous camera parameter features into the diffusion timestep embeddings with zero initialization.
• Replace by Conv: injecting parameters by replacing PAM with convolution-based layers that have the same number of parameters as PAM, but without spatial warping or channel-wise scaling.

As shown in Table R2‑4, both baselines underperform compared to PAM. TimeStep injection yields weaker parameter control due to the absence of localized spatial modeling. Conv injection provides partial improvements in parameter consistency but cannot explicitly capture geometric or global appearance variations.

By contrast, PAM, which combines spatial warping and channel‑wise scaling, achieves the best trade‑off between perceptual quality and parameter accuracy. These results confirm that our PAM design is critical for robust and precise parameter‑aware editing.

Table R2-4: Comparison of PAM vs. simpler baselines in aperture task.

L: Discussion of limitations and potential social impact in Section 6.

We will expand Section 6 in the revised version to provide a clearer discussion. Our work already covers a diverse set of camera parameters, and in future work we plan to extend the approach to more parameters and editing tasks. Moreover, we will also discuss the societal risks, especially wrongful image manipulation, and suggest controls like watermarking and post‑processing detection methods.

Reviewer ju7m,

We sincerely appreciate your feedback and have addressed your concerns accordingly. As the second review phase is nearing its end, we kindly ask if there are any remaining questions or clarifications we can assist with.

We sincerely thank you for your review. Below are your concerns and our corresponding responses:

W1: Discussion of related work such as ConceptSliders.

ConceptSliders is designed for common editing tasks such as changing human age and expressions or performing style transfer. However, it is not directly applicable to photorealistic image editing. However, our work proposes a photorealistic editing framework that supports tasks such as focus swapping, aperture adjustment, and exposure modification. We will include this discussion in the related work section.

W2 & Q1: Evaluation against ground-truth captures from professional digital cameras.

We evaluate CamEdit on three publicly available datasets with ground-truth captures from digital cameras: VABD [1] for aperture, EBB Val [2] for focal plane, and SICE [3] for shutter speed.  As shown in Table R1‑1, CamEdit outperforms both fine-tuned image editing methods (UltraEdit*, retrained on CamEdit50K) and rendering-based approaches (DrBokeh, CLODE) across all tasks, achieving superior SSIM and LPIPS scores.

Table R1-1: Quantitative comparison of CamEdit on three real datasets (VABD for Aperture, EBB for Focal Plane, and SICE for Shutter Speed).

[1] Andrey Ignatov et al. AIM 2020 Challenge on Rendering Realistic Bokeh. ECCVW, 2020.   [2] Kang Chen et al. Variable Aperture Bokeh Rendering via Customized Focal Plane Guidance. arXiv, 2024. [3] Jianrui Cai et al. Learning a Deep Single Image Contrast Enhancer from Multi-Exposure Images. TIP, 2018.

W3&Q3: Advantages of text-based input vs. other controls.

• Embedding camera parameters as textual inputs leverages the T2I diffusion model’s prior knowledge, improving the model’s understanding of the parameters and performing more accurate camera parameter editing, as demonstrated in Section 5.3 and Figure 7 of the main paper, compared to direct parameter mapping or linear approaches.    
• Moreover, the digital parameters of textual inputs can be directly mapped to slider‑based or other controls, with the editing software’s UI adapting to the user’s choice.

W4: Concern about editing only one parameter at a time.

We conduct an experiment where the model is trained to adjust all three parameters simultaneously. In this experiment, we double the number of training parameters and include 5,000 synthetic images with multi-parameter variations. As shown in Table R1‑2, the unified model demonstrates accuracy and output quality comparable to single-parameter control, confirming its ability to effectively manage multi-parameter adjustments. We will incorporate these updated results and provide further clarification on this capability in the revised manuscript.

Table R1-2: Performance of CamEdit with single-parameter vs. joint multi-parameter control, evaluated across 200 images with various parameter settings.

W6: The user study with real human subjects.

We conducted a user study with 15 participants who have expertise in photography. Each participant evaluated 20 image sets per task across all three tasks, rating photographic realism, parameter accuracy, and content preservation on a 0–4 scale. As shown in Table R1‑3, CamEdit consistently achieves the highest scores across all metrics, confirming the validity of our results under human evaluation.

Table R1-3: User Study Scores (0–4 scale) on Photographic Realism, Parameter Accuracy, and Content Preservation

W5&Q1: Concern about the lack of motion blur modeling for shutter speed and omission of lens artifacts (e.g., distortions).  

1. On motion blur and shutter speed:

• CamEdit models shutter speed via exposure adjustment, following standard single-image shutter modeling practices (e.g., PhotoGen [4]), which are widely used in computational photography.
• Motion-induced effects like blur from object or camera movement require multi-frame data and are beyond the scope of this single-image editing framework.
• The slightly lower GPT‑4o score is mainly due to brightness changes, as the model tends to penalize low-light images. Nonetheless, as shown in Table R3‑2 and R3‑3, CamEdit consistently outperforms existing methods in both human studies and comparisons with real ground-truth data.

2. On lens artifacts, parameter scales, and real-camera comparisons: Complex lens artifacts such as chromatic aberrations and geometric distortions are not controlled by camera parameters, and are therefore outside the scope of this work.

[4] Yu Yuan et al. Generative photography: Scene-consistent camera control for realistic text-to-image synthesis. CVPR, 2025.

W7: Unequal effectiveness of camera parameter adjustments, as some destructive effects cannot be undone.

As noted in Lines 34–36 and 89–94, CamEdit is consistent with existing rendering-based methods and physical camera models, and is designed for precise camera parameter adjustments. When input image details are lost, reversibility is not feasible for any current editing framework, as the missing information cannot be recovered. Addressing such cases would require dedicated image restoration strategies, which we identify as a promising direction for future research.

Q2&Q4: Clarifying the underlying camera model and data pipeline.

In the revised version, we will integrate the camera model into the main method section for a unified explanation of all referenced effects, and move the data generation process and image sources from the supplementary into the main text for clarity.

Q5: Consistency of generations

We evaluate the consistency of CamEdit across different parameter settings using 200 images per task. For each image, we compute the mean and standard deviation of the DINO similarity between the generated outputs and their corresponding inputs, across five different parameter settings for the same image.

CamEdit demonstrates high consistency with low variation:

Aperture: 0.83 ± 0.09  

Focal Plane: 0.84 ± 0.11

Shutter Speed: 0.93 ± 0.06  

For aperture and focal plane, we further analyzed foreground and in-focus regions separately, where the DINO standard deviations drop to 0.04 and 0.05, respectively. These results indicate that CamEdit’s edits are stable, with negligible changes outside the parameter-driven effects.

Qualitative results (Figure 1, Figure 5 in the main paper; Figure S3, Figure S5 in the supplementary) also show that CamEdit modifies only the target camera attributes without affecting unrelated scene content, further confirming the consistency of CamEdit.

L: The societal impact of photographic manipulations.

In the revised version, we will explicitly discuss potential societal impacts, including the risk of misuse for deceptive image manipulation, and outline mitigation strategies such as watermarking and edit-detection mechanisms.

About Typos.

We will correct all typos in the revision.

Dear Reviewer rxHt,

We sincerely appreciate your feedback and have addressed your concerns accordingly. As the second review phase is coming to a close, we kindly ask if there are any remaining questions or clarifications we can assist with.