InFlux++ Synth: Dynamic Intrinsics Dataset

• InFlux++ Synth is a procedurally generated video dataset with precise per-frame annotations for dynamic camera intrinsics.
• It leverages bounded random walks and Bézier interpolation to simulate smooth, realistic variations in focal length and distortion.
• Integrated with real-world benchmarks, the dataset improves focal-length estimation by about 6 percentage points for RGB-based calibration models.

InFlux++ Synth is a large-scale procedurally generated synthetic video dataset for estimating dynamic camera intrinsics from RGB video, introduced as one component of InFlux++, alongside the InFlux++ Real benchmark (Liang et al., 6 Jul 2026). It was designed to address two obstacles identified for dynamic-intrinsics estimation: scarce training data with limited intrinsics diversity, and benchmarks with limited scene and camera-motion diversity. InFlux++ Synth contains 1,841 high-resolution videos and 441,840 annotated frames, with accurate per-frame ground-truth intrinsics for training dynamic intrinsics prediction models; a subset also includes per-frame pose, depth, and normals (Liang et al., 6 Jul 2026).

1. Position within dynamic-intrinsics research

Camera intrinsics are vital for recovering 3D structure from 2D video, yet most 3D algorithms assume fixed intrinsics throughout a video. The InFlux++ work frames this assumption as frequently invalid for real-world in-the-wild videos, making per-frame intrinsics estimation from RGB images a critical problem for robust 3D reconstruction and related video-based geometry pipelines (Liang et al., 6 Jul 2026).

InFlux previously established the first real-world benchmark with per-frame ground-truth intrinsics for dynamic-intrinsics videos. InFlux++ extends that direction with two complementary components: InFlux++ Synth, which provides synthetic training data, and InFlux++ Real, which broadens real-world evaluation. Within that division of labor, InFlux++ Synth is explicitly intended to supply large-scale supervision with substantial intrinsics diversity, including changes in camera zoom and focus, dynamic objects, and realistic rendering effects such as lens distortion and defocus blur (Liang et al., 6 Jul 2026).

A common misconception is to treat dynamic intrinsics as equivalent to focal-length variation alone. The dataset description does not support that simplification. It reports per-frame intrinsics matrix $K$, per-frame camera focal length (CFL), lens focal length (LFL), lens-to-object distance (LTO), and supports radial and optionally tangential distortion. This indicates that the benchmarked problem includes time-varying optical state rather than only a scalar zoom trajectory.

2. Dataset composition and annotation schema

InFlux++ Synth contains 1,841 procedurally generated videos and 441,840 frames. Each video is 240 frames long at 24 FPS, corresponding to 10 seconds per video. Frames are distributed as RGB 8-bit PNG images at resolution $1280 \times 720$ pixels and frame rate 24 FPS (Liang et al., 6 Jul 2026).

Every frame includes the following annotations: the intrinsics matrix $K$, per-frame camera focal length (CFL), lens focal length (LFL), lens-to-object distance (LTO), and camera-to-world pose $T$ as a $4 \times 4$ matrix. A subset of 818 videos, comprising 196,320 frames, additionally includes per-pixel depth maps in meters and surface normal maps, each provided in both defocus-blurred and “pinhole” sharp versions (Liang et al., 6 Jul 2026).

Scene Type	Videos	Frames
Indoor	1,040	249,600
Nature	801	192,240
Total	1,841	441,840

The depth-and-normal subset spans 818 videos and 196,320 frames (Liang et al., 6 Jul 2026). The procedural scene coverage includes living rooms, kitchens, offices, caves, forests, deserts, and underwater environments. This breadth is relevant because the abstract identifies limited scene diversity as a major weakness of prior benchmarks. A plausible implication is that InFlux++ Synth is intended not merely as an intrinsics dataset, but as a controlled stress-test for RGB-based intrinsics estimators under diverse geometric and photometric conditions.

3. Camera intrinsics formulation and physical optics model

The dataset reports intrinsics in the standard zero-skew form

$$K = \begin{pmatrix} f_x & 0 & c_x \ 0 & f_y & c_y \ 0 & 0 & 1 \end{pmatrix},$$

where $f_x, f_y$ are focal lengths in pixels and $(c_x, c_y)$ is the principal point; skew is set to zero (Liang et al., 6 Jul 2026).

Zoom and focus changes are synthesized by sampling the physical lens focal length (LFL) and lens-to-object distance (LTO), then computing the effective camera focal length (CFL) through the thin-lens equation,

$$\frac{1}{\mathrm{LFL}} = \frac{1}{\mathrm{CFL}} + \frac{1}{\mathrm{LTO}}, \qquad \mathrm{CFL} = \frac{\mathrm{LFL}\,\mathrm{LTO}}{\mathrm{LTO}-\mathrm{LFL}}.$$

The dataset description states that this automatically induces realistic lens breathing, namely field-of-view shifts when LTO varies (Liang et al., 6 Jul 2026).

Temporal variation in intrinsics is generated through bounded random walks. The initial lens focal length is sampled as $$\mathrm{LFL}_0 \sim \mathrm{Uniform}(8\,\mathrm{mm}, 100\,\mathrm{mm})$$, and subsequent keyframes are updated by $1280 \times 720$0 while remaining within $1280 \times 720$1 mm. LTO is parameterized by $1280 \times 720$2 over the visible depth interval $1280 \times 720$3, and $1280 \times 720$4 follows a similar bounded walk. Keyframes are connected through Bézier interpolation to ensure smooth variation (Liang et al., 6 Jul 2026).

Distortion follows the Brown–Conrady model:

$1280 \times 720$5

$1280 \times 720$6

where $1280 \times 720$7 are normalized coordinates, $1280 \times 720$8, and $1280 \times 720$9 are distortion coefficients (Liang et al., 6 Jul 2026). In the default Synth loader for training “model-agnostic” methods, only $K$0 are sampled in order to match the two-parameter distortion heads of most calibration networks; tangential $K$1 can be enabled optionally.

The reported per-frame statistics define the dataset’s operating range. CFL, defined as $K$2, spans roughly 8 mm to 100 mm, corresponding in pixels to approximately 300 px to 4,000 px. LTO spans approximately 0.04 m to 12 m indoors and 0.7 m to $K$3 in nature scenes. The distortion parameter $K$4 is centered near 0 with 99% in $K$5, while $K$6 is near zero with tails clipped to $K$7. Principal-point offsets $K$8 are negligible in pure renders but become wide-ranged after the distortion loader, reaching $K$9 px (Liang et al., 6 Jul 2026).

4. Scene generation, camera trajectories, and rendering effects

The procedural scenes are built on Infinigen and Infinigen Indoors using Blender Cycles, which automatically places geometry, materials, lights, and dynamic actors such as animals, falling leaves, and water (Liang et al., 6 Jul 2026). This procedural basis is central to the dataset’s stated goal of overcoming scarce real training data and limited intrinsics diversity.

Camera motion is generated with an RRT-based trajectory planner combined with Bézier interpolation. The pipeline adds view validation that rejects views with more than 65% dominated by one surface normal, with the stated aim of ensuring rich parallax and geometry (Liang et al., 6 Jul 2026). This is significant because dynamic-intrinsics estimation can be confounded by scene degeneracy; the rejection criterion suggests that the dataset construction attempts to avoid trajectories with weak geometric cues.

Several optical effects are explicitly incorporated. Defocus blur is rendered using Blender’s thin-lens depth of field with a finite aperture. Motion blur is generated inherently by Cycles when the camera or objects move between sub-samples. Lens distortion is applied offline in the data loader for on-the-fly augmentation via the Brown–Conrady model. Color-photometric augmentations include gamma, tone curves, brightness/contrast, jitter, JPEG noise, and ISO noise, following AnyCalib settings, but with no additional down-sampling or sharpening (Liang et al., 6 Jul 2026).

A plausible implication is that InFlux++ Synth is structured to preserve blur and distortion cues as first-class supervision signals rather than nuisance artifacts. The recommendation to avoid extra down-sampling or sharpening reinforces that interpretation.

5. Parameter diversity and statistical profile

The dataset description reports intrinsics histograms with a uni-modal CFL spread from approximately 400 px to 2,000 px, LTO distributions concentrated around $T$0 m indoors and $T$1 m in nature scenes, and $T$2 values tightly centered around zero with realistic tails (Liang et al., 6 Jul 2026). These statistics complement the broader range summary and indicate that the dataset is not merely covering extreme values, but also concentrating substantial mass in mid-range operating regimes.

One example of dynamic trajectories is given for a typical indoor video: LFL keyframes of 12 mm $T$3 20 mm $T$4 45 mm over 240 frames, LTO keyframes of 1.2 m $T$5 0.8 m $T$6 2.5 m, and a resulting CFL time series described as smooth and physically plausible (Liang et al., 6 Jul 2026). Because keyframes are connected by Bézier interpolation and constrained by a bounded random walk, the temporal process is designed to avoid abrupt intrinsics discontinuities.

This suggests that the dataset targets temporal estimators as well as frame-wise predictors. The source text makes that intention explicit in its usage recommendations, which note that the random-walk intrinsics ensure that a network sees smooth temporal variation. That design choice distinguishes InFlux++ Synth from datasets where per-frame variation may be independent and therefore less representative of real camera operation.

6. Training integration, recommended usage, and observed effects

The released PyTorch data loader is intended to fetch images, intrinsics, pose, and distortion on the fly (Liang et al., 6 Jul 2026). The usage recommendations specify two training modes aligned with model class: feed entire clips to temporal models, or sample frames uniformly for frame-wise predictors. This is one of the clearest statements of the benchmark’s expected use pattern.

For augmentation, the recommendation is to follow AnyCalib’s “$T$7” augmentation recipe while disabling geometric down-sampling and sharpening in order to preserve synthetic blur cues. When training networks that predict only two radial distortion parameters, the loader should freeze $T$8. For geometric cropping and resizing, the recommended aspect-ratio sampling range is $T$9, with the stated rationale of avoiding discarding corner pixels, which is important for distortion learning (Liang et al., 6 Jul 2026).

The empirical impact reported for training is narrowly defined but concrete: fine-tuning AnyCalib on InFlux++ Synth for 15 epochs improved focal-length recall@10% by approximately 6 percentage points on InFlux++ Real and by approximately 6 percentage points on InFlux (Liang et al., 6 Jul 2026). At the same time, a small drop in principal-point and EPE recall is reported, and the source attributes this to a possible mismatch between training losses and projection-error metrics. This is an important corrective to overly broad claims about synthetic supervision. The evidence presented supports consistent gains for focal-length estimation, but not uniform improvement across every calibration metric.

That result also addresses a common concern about synthetic data for camera calibration. The abstract states that fine-tuning existing intrinsics prediction methods on InFlux++ Synth consistently improves focal-length estimation across both InFlux++ Real and InFlux, suggesting that synthetic supervision is promising for RGB-based intrinsics prediction (Liang et al., 6 Jul 2026). The wording “promising” is appropriately limited: it indicates positive transfer without claiming that synthetic training fully resolves evaluation on real videos.

7. Relation to the broader InFlux++ benchmark

InFlux++ consists of two components with distinct functions. InFlux++ Synth supplies large-scale synthetic supervision, while InFlux++ Real extends the original InFlux benchmark with 514K+ newly captured frames across 334 high-resolution videos, spanning a wider range of scenes and camera motions (Liang et al., 6 Jul 2026). This pairing is central to the project’s methodology: synthetic data addresses training scarcity, and real data addresses evaluation realism.

Within that framework, InFlux++ Synth can be understood as the training-oriented half of a benchmark suite for dynamic intrinsics. It provides fully accurate per-frame intrinsics, including distortion, camera pose, and, for 818 videos, ground-truth depth and normals (Liang et al., 6 Jul 2026). In contrast, the real benchmark expands scene and motion diversity for evaluation. This division suggests a deliberate separation between supervision scalability and deployment realism.

The dataset, benchmark, code, videos, submission instructions, and live leaderboard are hosted at the project website, indicating that InFlux++ is intended as an ongoing benchmarked resource rather than a static release (Liang et al., 6 Jul 2026). A plausible implication is that InFlux++ Synth is likely to be used not only for standalone model training, but also for pretraining, fine-tuning, and ablation studies on temporal intrinsics estimators whose final assessment is carried out on the real benchmark.