Blur Synthesis Pipeline Techniques

• Blur synthesis pipeline is a structured methodology that algorithmically generates blurred images using physical principles and parametric flexibility.
• It simulates blur via raw-domain averaging combined with accurate ISP rendering to faithfully mimic sensor characteristics and nonlinear effects.
• Differentiable blur operators and latent space interpolation enable end-to-end deep learning optimization, enhancing deblurring and restoration performance.

A blur synthesis pipeline is a structured methodology for algorithmically generating blurred images, often as intermediate supervision, data augmentation, or training targets for tasks such as deblurring, denoising, novel view generation, and generative model control. Contemporary pipelines emphasize physical accuracy (sensor modeling, optical integration, hardware-specific artifact simulation), parametric flexibility (spatial and temporal blur profiles), and differentiability for enabling end-to-end optimization in modern deep networks.

1. Physical Principles and Mathematical Foundations

The physical origin of image blur is the temporal and/or spatial integration of scene radiance on a sensor, modulated by relative motion (camera shake, object movement), lens geometry (defocus via finite aperture), or hardware mechanics (rolling shutter, jitter). A precise description requires:

• Motion Blur: Sensor integration over moving scene/camera, modeled as

$B(x, y) = \int_{0}^\tau S(x, y, t) \, dt$

where $S(x, y, t)$ is the instantaneous sharp radiance and $\tau$ is the exposure duration (Cao et al., 2022, Wei et al., 2022, Lee et al., 2023).

• Defocus Blur: Finite-aperture optics induce a point-dependent convolution (circle of confusion, CoC). The per-pixel radius under a thin-lens model is

$$C_p = \frac{|d_p - f_d|}{d_p} \cdot \frac{f^2}{N(f_s f_d - f)}$$

with per-pixel depth $d_p$, predicted focus distance $f_d$, focal length $f$, f-number $N$, and learned scale $f_s$ (Shrivastava et al., 7 Oct 2025, Wang et al., 27 May 2024, Morris et al., 1 Jul 2024).

• Rolling Shutter & Jitter: Sequential sensor readout or platform oscillations introduce non-uniform, position-dependent blurring modelled via composite warping and temporal averaging (Seiskari et al., 20 Mar 2024, Chen et al., 16 Jan 2024).

2. Modern Algorithmic Pipelines: Raw-Domain Synthesis and ISP Modeling

State-of-the-art blur pipelines fundamentally distinguish themselves along (i) the image domain in which blur is simulated, and (ii) explicit simulation of the camera’s image signal processor (ISP):

• Raw-Domain Averaging: Blurring is synthesized by averaging high-frame-rate RAW frames. This simulates sensor exposure linearly, preserving physical integration fidelity and circumventing nonlinear camera response artifacts (Cao et al., 2022, Wei et al., 2022). Mathematically:

$$B_\mathrm{raw}(x, y) = \frac{1}{\tau} \sum_{i=0}^{\tau - 1} S_\mathrm{raw}[i](x, y) + N(0, \sigma^2(x, y))$$

where noise is injected according to real sensor statistics.

• ISP Rendering: The synthesized blur in RAW space is rendered to RGB using an accurate or learned ISP—either via software pipelines (e.g., RawPy, DarkTable) or deep models (e.g., CycleISP)—to mimic real camera color, gamma, demosaicing and noise properties (Cao et al., 2022, Wei et al., 2022). This step is essential to avoid domain gaps due to differences in color and nonlinearity.
• Parameterization of Blur Strength: The duty cycle ($\tau/T$), temporal sampling, and per-frame exposure drive the blur magnitude. Increasing $\tau$ proportionally increases the blur support, directly controlling the synthetic effect magnitude.

3. Learning, Transfer, and Differentiable Blur Operators

Recent frameworks employ learned, fully differentiable modules to encode and synthesize complex, real-world blur:

• Kernel Space Encoding: Datasets of sharp-blur image pairs are compressed into a low-dimensional latent blur-kernel space via an encoder–decoder setup, enabling transfer of empirically observed blur to new domains, or faithful data augmentation. The mapping

$$\text{blur code}\; k = G_\phi(x, y),\quad \text{synthesize:}\; \hat{y} = F_\theta(x, k)$$

can generalize to unseen blur and is compatible with end-to-end training of deblurring architectures (Tran et al., 2021).

• Latent-Space Interpolation & Extrapolation: For defocus, linearly interpolating or extrapolating between latent codes (autoencoder embeddings) produces continuous blur/deblur manipulation. Given latent codes $z_a, z_c$ at two known blur levels, intermediate ($z_b$) or extrapolated sharp representations are synthesized as

$$z_b' = \alpha z_a + (1-\alpha) z_c,\quad \tilde{x}_b' = D(z_b')$$

encouraging linearity in latent space via implicit or explicit regularization (Mazilu et al., 2023).

• End-to-End Blur Synthesis in Generative Models: Fine-grained defocus control in diffusion models is achieved by differentiable propagation through all-in-focus image synthesis, monocular depth prediction, learned focus-map aggregation, and a physically-correct, local disk convolution. This maintains gradient flow through all stages, making lens blur controllable via EXIF/aperture metadata and inferrable during unsupervised training (Shrivastava et al., 7 Oct 2025).

4. Domain-Specific Blur Synthesis: 3D Scenes, Video, and Specialized Hardware

Pipelines are adapted to specific domain demands:

• 3D Scene/Novel View Synthesis: Gaussian Splatting (3DGS) with finite-aperture models allows explicit, differentiable defocus (thin-lens, variable CoC) for post-capture refocusing and blur removal. Each splat’s screen-space kernel is convolved with a CoC-adaptive Gaussian, and layered via alpha compositing (Wang et al., 27 May 2024). For dynamic scenes, motion and rolling shutter are incorporated by integrating projected splat centers over pose trajectory and temporal exposure (Seiskari et al., 20 Mar 2024).
• Extreme Motion Blur in View Synthesis: ExBluRF decouples motion-blur formation as time-integration along 6-DOF Bézier-parametrized camera trajectories and leverages efficient voxel-grid radiance fields, giving

$$B(p) = \int_{t_0}^{t_1} \mathcal{V}(\mathbf{r}_t(p)) dt \approx \frac{1}{N} \sum_{i=1}^N \mathcal{V}(\mathbf{r}_{t_i}(p))$$

for pixel $p$ (Lee et al., 2023).

• Remote Sensing/Pushbroom Sensors: The CDSM-based pipeline explicitly simulates geometric distortion and sub-pixel blur due to platform jitter via time-varying warps, determined by a sum-of-sinusoids model and bilinear warping of the ground-truth scene (Chen et al., 16 Jan 2024).

5. Advanced Blur Types: Multi-View, Focal, and Spatially-Variant Blur

Emerging pipelines synthesize complex, spatially-varying blur effects:

• Focal Blur with Depth Guidance: Video focal-blur synthesis relies on depth estimation, stochastic key-point focal distance parameterization, and per-pixel spatially variant Gaussian convolution. For video, the focal plane can move within a sequence, enabling precise training supervision via generated blur maps and depth maps (Morris et al., 1 Jul 2024).
• Multi-view Bokeh and Directional PSFs: For synthetic DoF effects and “bokeh” rendering, rotated dual-pixel PSFs (half-disk) are aligned with depth layers, blurred, and composited in a back-to-front order. Multi-view “aperture slices” are synthesized by rotating these PSFs, supporting NIMAT-style motion and parallax (Abuolaim et al., 2021).
• Motion Blur from Sharp Image Pairs: Differentiable line-prediction layers, combined with U-Net feature supervision, allow networks to learn per-pixel motion line trajectories and sample weights. This architecture synthesizes motion blur from two sharp input frames—the synthetic output closely matches physically averaged (240 fps) ground-truth (Brooks et al., 2018).

6. Quantitative Impact and Guidelines for Evaluation

The accuracy of blur-synthetic data directly determines the generalization of deblurring and restoration models to real-world cases:

• PSNR/SSIM Gains: Pipelines that synthesize blur in RAW space and apply camera-matched ISPs consistently yield 0.5–5 dB PSNR improvements over RGB-domain, kernel-based, or low-fps-only pipelines on real-world test sets (Cao et al., 2022, Wei et al., 2022).
• Ablation Protocols: Evaluations compare RAW-based vs RGB-based blur, parametric vs learned ISPs, spatially invariant vs depth-variant PSFs, and autoencoder regularization strategies for synthetic data (Wei et al., 2022, Mazilu et al., 2023).
• Synthetic-to-Real Domain Bridging: Use of real-capture statistics (frame rate, exposure, noise), ISP-consistent color handling, and parameter randomization (duty cycle, kernel size) are critical for producing blur that enables networks to generalize to real camera data.

Pipeline Domain	Blur Formation Domain	Key Features/Components
Video/Motion	RAW + ISP	Frame averaging, Poisson noise, camera ISP modeling
Focal/Defocus	RGB + Depth	Depth maps, spatially varying Gaussian PSF
3D Scene Synthesis	Gaussian splats	Thin-lens, CoC modeling, differentiable convolution
Blur Latent Autoenc.	Latent space	Linearity regularization, interpolation/extrapolation
Pushbroom Jitter	Geometric warp	Sum-of-sinusoids, subinterval averaging

7. Open Challenges and Future Perspectives

The field is advancing toward physically accurate, fully differentiable, domain-adaptive blur synthesis:

• Blurs involving nonrigid object motion, lighting time-varying artifacts, and real hardware distortions are only partially captured in current pipelines.
• Data-driven ISP synthesis and improved domain adaptation strategies are required to bridge camera-specific gaps for new deployment scenarios.
• Real-time hardware simulators and integration with generative diffusion backbones will push toward content-adaptive, controllable, and invertible blur effects suitable for interactive, photorealistic applications and robust model training.

Blur synthesis pipelines now form the backbone of data generation for deblurring, super-resolution, photorealistic rendering, and generative models, with physical realism and computational tractability as core design parameters (Cao et al., 2022, Wei et al., 2022, Wang et al., 27 May 2024, Morris et al., 1 Jul 2024, Shrivastava et al., 7 Oct 2025, Seiskari et al., 20 Mar 2024, Jonnalagadda et al., 12 Nov 2024, Brooks et al., 2018, Mazilu et al., 2023, Abuolaim et al., 2021, Tran et al., 2021, Lee et al., 2023, Park et al., 2020).