AirSketcher: Image-Driven Sketching & CFD

• AirSketcher is an image-driven pipeline that converts visual inputs into clean sketches or geometries for AR and CFD applications.
• The graphics variant employs a controllable diffusion model to denoise noisy hand-tracking images and produce coherent sketches without specialized hardware.
• In engineering, AirSketcher implements CFD workflows that extract geometry from side-profile images to facilitate rapid dune analysis and industrial digital twin diagnostics.

The name AirSketcher appears in recent arXiv literature in multiple, technically unrelated contexts. In graphics and AR-adjacent research, the closely related term AirSketch denotes marker-less air drawing: generating a clean, aesthetically pleasing sketch from a video of a person drawing in the air with their finger, using only ordinary camera input and hand tracking. In computational engineering, AirSketcher denotes a lightweight image-to-geometry CFD workflow and digital-twin engine that starts from a side-profile image or schematic, auto-extracts geometry, solves a 2D incompressible RANS problem, and reports interpretable transport or energy metrics. The term therefore spans generative sketch reconstruction, dune-flow analysis, and industrial ventilation maintenance modeling in the cited corpus (Lim et al., 2024, Pattanapol et al., 23 Sep 2025, Pattanapol et al., 31 Jan 2026).

1. Terminological scope and principal usages

A recurring source of confusion is nomenclature. The graphics paper introduces AirSketch as the term for marker-less air drawing, but its detailed description explicitly groups the topic under “AirSketch / AirSketcher.” By contrast, the engineering papers use AirSketcher as the name of a CFD workflow or solver. These usages are linked by an image-first workflow—both begin from visual input rather than specialized hardware or full 3D models—but they address different technical problems, use different mathematical machinery, and produce different outputs (Lim et al., 2024, Pattanapol et al., 23 Sep 2025, Pattanapol et al., 31 Jan 2026).

Usage in the cited literature	Domain	Core function
AirSketch	Generative graphics, AR/VR	Translate noisy hand-tracking images into clean sketches
AirSketcher	Coastal dune CFD	Extract a dune outline from a side-profile image and compute near-surface flow and transport proxies
AirSketcher	Industrial digital twin	Model clogged-filter fan-room aerodynamics, pressure loss, airflow, and maintenance ROI

The graphics usage centers on human motion as a noisy proxy for intended drawing. The engineering usages center on side-profile geometry as a proxy for environmental or facility-scale flow domains. A plausible implication is that “AirSketcher” functions less as a single canonical platform name than as a label reused for compact, image-driven modeling pipelines.

2. AirSketch as marker-less generative motion-to-sketch

In the 2024 graphics formulation, AirSketch addresses the problem of creating sketches from hand motions without expensive AR/VR headsets, controllers, or physical markers. The task is framed as Generative Motion to Sketch: infer the intended clean sketch from a very noisy motion trace. The motivating claim is that hand-tracking output is typically noisy, distorted, and incomplete, so simply rendering the tracked trajectory yields sketches that are visually poor or unrecognizable. The proposed remedy is a controllable diffusion model trained to translate a noisy hand-tracking image into a clean sketch while preserving the essential visual cues embedded in the motion data (Lim et al., 2024).

The method uses ControlNet-style spatial conditioning. Rather than conditioning on a clean edge map, depth map, or pose map, it conditions on a corrupted sketch-like raster representation derived from hand tracking. The model input consists of a noisy raster sketch image together with a text prompt of the form $\text{“a black and white sketch of a <category>”}$, and the output is a clean sketch image. This image-space formulation is presented as an alternative to coordinate-sequence auto-regression such as Sketch-RNN. The stated advantages are constant computation with respect to sketch length, order invariance, and compatibility with pretrained image diffusion models (Lim et al., 2024).

The core pipeline is straightforward. A ground-truth sketch $\mathbf{x}_0$ is corrupted by an augmentation operator $\mathcal{A}(\cdot)$ designed to mimic air-drawing and tracking artifacts; the corrupted image is then supplied as the control condition to a diffusion model trained to reconstruct the clean target. At inference time, the synthetic corruption is replaced by a real tracking image extracted from a hand-motion video using a hand detector or tracker such as MediaPipe Hands. This establishes the central interpretation of AirSketch: the model is not merely generating an image from text, but denoising a structured, highly unreliable motion-derived signal into a coherent sketch.

3. Self-supervised learning formulation, augmentation design, and datasets

The central technical obstacle is the lack of paired real data. There is no large dataset of real air-drawing tracking images matched to clean target sketches, so the method adopts a self-supervised augmentation-based training strategy. Clean sketches are drawn from Quick, Draw!, corrupted with augmentations that imitate tracking artifacts, and then reconstructed. The target remains the original clean sketch, eliminating the need for manual pair annotation (Lim et al., 2024).

The paper starts from the standard conditional diffusion denoising objective

$$\mathcal{L} = \mathbb{E}_{\mathbf{x}_0, \mathbf{t}, \mathbf{c}_t, \mathbf{c}_f, \epsilon \sim \mathcal{N}(0,1)} \left[ \left\| \epsilon - \epsilon_\theta(\mathbf{x}_t, \mathbf{t}, \mathbf{c}_t, \mathbf{c}_f) \right\|_2^2 \right]$$

and replaces the spatial condition with the corrupted sketch image,

$$\mathcal{L} = \mathbb{E}_{\mathbf{x}_0, \mathbf{t}, \mathbf{c}_t, \mathbf{c}_f, \epsilon \sim \mathcal{N}(0,1)} \left[ \left\| \epsilon - \epsilon_\theta(\mathbf{x}_t, \mathbf{t}, \mathbf{c}_t, \mathcal{A}(\mathbf{x}_0)) \right\|_2^2 \right].$$

Here $\mathbf{x}_0$ is the clean sketch, $\mathbf{x}_t$ is the noised latent or image at diffusion timestep $t$, $\epsilon$ is Gaussian noise, $\epsilon_\theta$ is the UNet denoiser, and $\mathbf{x}_0$0 is the text condition. The conceptual adaptation is important: the control branch is trained to interpret a distorted sketch-like condition rather than a clean, spatially exact constraint.

The augmentation design is divided into three groups. Local augmentations model fine-grained hand noise through stroke distortion via random wave functions, random sharp or smooth spikes, and jitter perturbations. Structural augmentations model global drawing issues through distortion of overall sketch shape or aspect ratio, repositioning on the canvas, and random resizing of strokes. False strokes model unintended extra strokes such as transitional lines between strokes and random extra lines. Some augmentations are applied with a 50% chance, and misplacement and resize are mutually exclusive because together they can destroy the sketch too much. The underlying rationale is that corrupted inputs should preserve enough evidence of the intended object for the model to learn which features are meaningful shape cues and which are artifacts (Lim et al., 2024).

Two air-drawing datasets are introduced because no prior dataset pairs air-drawing motion with ground-truth sketches. The Synthetic Air-Drawing Dataset is constructed by taking clean Quick, Draw! sketches as ground truth, animating a 3D arm in Unity using inverse kinematics to trace the stroke sequence, and recording videos of the hand drawing in air. It contains 50 categories, 100 sketches per category, for a total of 5000 sketch-video pairs, recorded at 1920×1080, 60 FPS. The Real Air-Drawing Dataset consists of a human user attempting to replicate Quick, Draw! sketches in the air, recorded at 1280×720, 30 FPS, with 10 samples per category from the same 50 categories, totaling 500 video-sketch pairs. To test generalization, the paper holds out ten categories: $\mathbf{x}_0$1 Training uses 60K Quick, Draw! sketches, keeping only the top 5% most semantically aligned sketches per category by CLIP image-text similarity, and excluding the held-out categories. The primary backbone is Stable Diffusion XL (SDXL); Stable Diffusion 1.5 is also compared; LoRA fine-tuning is used to give the backbone the basic ability to generate sketches in the desired style (Lim et al., 2024).

4. Empirical behavior, prompting effects, and limitations of the generative system

Evaluation uses SSIM, Chamfer Distance (CD), LPIPS, CLIP I2I, and CLIP I2T, combining local structural similarity, global shape distance, perceptual similarity, and semantic recognizability. The central experimental finding is that augmentation-based training works: without augmentations, ControlNet fails to improve over the noisy tracking input, whereas with augmentations the model removes noise, preserves object shape, generates visually pleasing sketches, and remains robust on both synthetic and real tracking inputs (Lim et al., 2024).

The reported gains are consistent across synthetic and real data, and across seen and unseen categories. On the real dataset with SDXL, compared to the tracking input, SSIM improves by about 10%, LPIPS decreases by about 6%, and CD decreases by about 21%. For real seen categories, the tracking baseline is SSIM 0.55, CD 32.36, LPIPS 0.42, whereas the generated output with augmentation plus SDXL reaches SSIM 0.64, CD 25.46, LPIPS 0.36. For real unseen categories, tracking yields SSIM 0.54, CD 33.92, LPIPS 0.42, while generation reaches SSIM 0.63, CD 24.26, LPIPS 0.38. Ablation further shows that local augmentations are most important for local detail and jitter removal, while false strokes and structural augmentations improve global shape and reduce false-line following. Using all three augmentation groups produces SSIM 0.62, CD 27.33, LPIPS 0.37, CLIP I2I 0.84, and CLIP I2T 0.29 (Lim et al., 2024).

The model also generalizes to held-out classes, which the authors interpret as evidence for a category-agnostic transformation from noisy sketch-like input to clean sketch rather than simple category memorization. Text conditioning is materially important. When inputs are heavily corrupted or categories are unseen, prompts become more important; without prompts, performance drops notably and the model may hallucinate a plausible seen category instead of the correct one. A specific qualitative failure reported is that an unseen “snail” may be generated as a seen category such as a “fan.” This indicates that the model balances visual evidence from the motion trace against semantic guidance from text. The paper also reports sketch completion from partially missing drawings, text-conditioned stroke styling such as color and thickness, and better fidelity from ControlNet than T2IAdapter in some qualitative examples (Lim et al., 2024).

The stated limitations are narrow but consequential. The evaluation is mainly on Quick, Draw!-style drawings with few strokes, so the method is restricted to simple/cartoon-like sketches. It does not explore complex color artwork. It remains dependent on text prompts for ambiguous or unseen cases, and it still depends on hand-tracking quality even though the model is designed to tolerate noisy MediaPipe outputs. The authors position the method as an initial step toward marker-less air drawing rather than a complete solution for general-purpose artistic generation.

5. AirSketcher as a field-anchored CFD workflow for dune-shape sensitivity

In coastal engineering, AirSketcher denotes a lightweight, field-anchored CFD workflow for rapidly testing how small dune-shape changes affect near-surface wind speed and, by extension, aeolian sand transport. The workflow begins with an imported side-on profile image or drawing, typically a PNG/JPG. It uses contour/edge detection to trace the dune outline, after which the user tidies the polyline with edit handles and calibrates scale using a known distance. Modified shapes are created by duplicating and editing the contour while keeping beach and backdune tie-points fixed, so that flow differences can be attributed to geometry rather than domain placement. The computational domain is then built directly from the surveyed profile on a uniform Cartesian grid in the 2D plane of the transect (Pattanapol et al., 23 Sep 2025).

The flow solver is a 2D incompressible Reynolds-averaged Navier–Stokes model with an eddy-viscosity closure and a projection method for pressure–velocity coupling. The governing continuity equation is

$\mathbf{x}_0$2

and the paper describes advection as explicitly advanced with variable time stepping, diffusion and turbulence terms as using second-order central differences, and convective terms as using a bounded upwind/central blend. Pressure correction is handled with a standard projection step. Turbulence is closed with the one-equation Spalart–Allmaras model, using canonical coefficients and no wall functions, so near-ground shear is resolved directly on the mesh. The inlet is a neutral atmospheric boundary layer using a power-law inlet with exponent $\mathbf{x}_0$3, applied consistently across scenarios, with a reference speed of about $\mathbf{x}_0$4 at $\mathbf{x}_0$5. The top boundary is free-slip/symmetry, the outlet uses zero-gradient conditions, the dune surface is no-slip, time stepping is constrained by a CFL criterion, and the solver monitors mass balance through inflow and outflow fluxes (Pattanapol et al., 23 Sep 2025).

The transport metric is deliberately simple and geometry-aware. AirSketcher draws a near-surface polyline along the transect, samples local speed $\mathbf{x}_0$6, and computes a Bagnold-type cubic line integral,

$\mathbf{x}_0$7

The paper also discusses a thresholded aeolian version involving a site-specific threshold speed $\mathbf{x}_0$8, but treats the thresholded and non-thresholded forms primarily as absolute-versus-relative variants; the ranking of geometries is stable because the cubic dependence heavily weights high-speed pockets (Pattanapol et al., 23 Sep 2025).

Field validation was conducted at Tomahawk Beach in Dunedin, New Zealand, along a steep foredune transect implicated in proposed sand-removal works. The transect was surveyed with a Leica total station and sampled at nine stations (ST1–ST9) spanning the beach approach, toe, stoss, crest, lee, and backdune. Each mobile mast carried five cup anemometers at nominal heights of 0.2, 0.5, 1.0, 2.0, and 5.0 m, plus a wind vane at 2.0 m; each stop lasted five minutes with logging every three seconds. A stationary reference mast near the stoss crest recorded concurrent wind speed and direction. Comparing simulated and measured vertical profiles at masts 1, 4, 8, and 9, the paper reports height-matched correlations of

$\mathbf{x}_0$9

The validation target was not only profile agreement but also reproduction of canonical dune aerodynamics: stoss-side speed-up, crest amplification, and lee-side recovery/decay, including a pronounced lee jet and later reattachment in the baseline case (Pattanapol et al., 23 Sep 2025).

Three geometries are analyzed: S1 baseline, the unmodified surveyed dune; S2 top-cut, where the crest is lowered over a finite crest window with smoothing on adjacent stoss and lee sections; and S3 back-cut, where the lee side is trimmed, shifting the crest landward and regrading the lee slope to the backdune tie-in. Using the validated setup and a consistent near-surface polyline, the integrated cubic proxies are

$\mathcal{A}(\cdot)$0

Normalized to baseline, these are 1.000, 0.830, and 0.603, implying reductions of about 17.0% for the top-cut and 39.7% for the back-cut. The ranking is $\mathcal{A}(\cdot)$1. The explanation given is consistent with hill-flow theory: crest flattening attenuates peak speed and weakens the lee jet, while lee-side truncation most strongly suppresses the lee jet and leads to the earliest reattachment. Because the transport proxy is cubic in speed, reducing high-speed pockets has an outsized effect. The paper further argues that in S3 the steep post-crest decay concentrates deposition near and just lee of the crest and within the cut, supporting “self-healing.” AirSketcher is therefore positioned as a rapid screening tool for dune-modification concepts rather than as a replacement for fully coupled morphodynamic modeling (Pattanapol et al., 23 Sep 2025).

6. AirSketcher as an industrial fan-room digital twin for maintenance optimization

A second engineering usage presents AirSketcher as a 2D, steady-state, incompressible Reynolds-Averaged Navier–Stokes solver for rapid digital-twin analysis of industrial ventilation systems. In this formulation, its capabilities include solving incompressible flow, using a structured Cartesian grid with automated local refinement, modeling turbulence with Spalart–Allmaras, representing filters as porous-media zones with a drag law tied to effective porosity, generating line-probe outputs for velocity and pressure, and producing an automated energy summary for ROI and maintenance decisions. The paper explicitly frames AirSketcher not merely as a CFD solver but as a physics-based digital twin for fan-room performance and filter-maintenance planning (Pattanapol et al., 31 Jan 2026).

The workflow is organized in two stages. The validation stage benchmarks the porous-media model against a published porous-screen wind-tunnel experiment from Politecnico di Milano. The virtual wind tunnel is 10 m × 2 m with a central porous region, and pressure drop is extracted along a centerline probe. At 10 m/s inlet velocity, the experimental target is 153 Pa and AirSketcher predicts 165 Pa, a deviation of +7.8%. At 20 m/s, the target is 612 Pa and AirSketcher predicts 650 Pa, a deviation of +6.2%. The intended physics check is the inertial pressure-drop law

$\mathcal{A}(\cdot)$2

The industrial application stage then applies the same solver to a draw-through fan-room configuration, inlet → filter bank → suction plenum → fan, in a 6 m × 4 m 2D room model (Pattanapol et al., 31 Jan 2026).

The governing equations are the incompressible Navier–Stokes equations,

$\mathcal{A}(\cdot)$3

$\mathcal{A}(\cdot)$4

For the porous filter, the paper uses an effective-porosity drag formulation,

$\mathcal{A}(\cdot)$5

and applies a semi-implicit damping term to enforce a Darcy–Forchheimer-type resistance law. The case study concerns a standard 50 kW draw-through fan room representative of industrial ventilation in settings such as glass manufacturing. The operating assumptions are a VFD-driven industrial fan system, 3,000 hours/year, an electricity tariff of $0.12/kWh, 80% porosity for the clean filter, and 50% porosity for the dirty filter, with the latter serving as a proxy for severe loading (Pattanapol et al., 31 Jan 2026).

The paper’s central interpretive claim is that the VFD masks filter clogging by increasing fan speed to maintain airflow. This control logic preserves visible process performance until the system approaches either its maximum available pressure head or its maximum speed limit, at which point the VFD enters a saturation regime and effective airflow collapses. Using AirSketcher, the study predicts 3,806 CFM for the clean filter and 1,831 CFM for the dirty filter, a loss of 1,975 CFM or about 52% reduction in airflow. The dirty case is described as exhibiting a severe static vacuum spike immediately downstream of the filter face, with fan effort increasingly spent on suction head rather than useful mass transport (Pattanapol et al., 31 Jan 2026).

The energy analysis uses an “upstream energy saving” logic based on ratios of flow and pressure effects, reporting

$\mathcal{A}(\cdot)$6

From a base energy of

$\mathcal{A}(\cdot)$7

the annual energy penalty avoided by restoring the clean baseline is given as 8,818 kWh, corresponding to $\mathcal{A}(\cdot)$8800, giving a payback estimate of

$\mathcal{A}(\cdot)$9

or roughly 9 months. The stated limitations are explicit: the model is 2D steady-state rather than fully 3D transient, porosity reduction is only a proxy for spatially non-uniform fouling, the clean and dirty porosities are simulation settings rather than measured microstructural properties, and the energy interpretation assumes unchanged fan efficiency and downstream resistance. Within those assumptions, AirSketcher is presented as a defensible, ROI-oriented maintenance-screening tool rather than a substitute for full motor-electrical or plantwide simulation (Pattanapol et al., 31 Jan 2026).

7. Relation to adjacent AR sketching systems

The broader AR sketching literature provides useful context for the graphics meaning of AirSketcher. RealityCanvas is a mobile AR sketching system for embedded and responsive scribble animations directly on live or recorded scenes. It does not solve the same problem as the diffusion-based AirSketch system, nor does it belong to the CFD line, but it extends the same general research concern with freehand, motion-responsive sketching beyond static placement into interactive animation (Xia et al., 2023).

RealityCanvas is designed around two goals: expressiveness and improvisation. Its design space is derived from an analysis of 172 existing scribble animation videos, filtered to 120 videos exhibiting embedded or responsive behavior. The resulting taxonomy comprises six common animation techniques: object binding, flip-book animation, action trigger, particle effects, motion trajectory, and contour highlight. The workflow is a three-step sequence: object selection and tracking, sketching, and assignment of an animation behavior. Tracking uses color tracking for objects and MediaPipe for bodies, hands, and face anchors; the body-tracking side leverages 33 body landmarks. The implementation is a web-based mobile AR prototype built with HTML Canvas, SVG, Konva.js, and Anime.js, using a standard RGB camera and supporting phones, tablets, and laptops (Xia et al., 2023).

The system’s technical emphasis differs from AirSketch. Instead of reconstructing a clean canonical sketch from noisy motion traces, RealityCanvas links sketches to trackers and updates them as real-world targets move or deform. The paper reports a lab study with 20 participants and expert interviews with 7 experts. Participants rated overall experience 5.65/7, ease of understanding 5.25/7, lower barrier to creating AR animation 4.95/7, flexibility 5.45/7, and expressive or high quality 5.3/7; average completion times were 3.25 min for a standardized task and 5.24 min for an exploration task. These findings suggest a neighboring design principle for AR sketching systems: low-cost mobile accessibility, direct sketch-to-animation coupling, and real-time responsiveness can be as important as generative reconstruction accuracy (Xia et al., 2023).

Taken together, the cited literature shows that AirSketcher is best understood as a name attached to several compact, image-driven modeling paradigms. In graphics, it refers to marker-less sketch reconstruction from noisy hand motion; in coastal and industrial engineering, it refers to 2D CFD workflows that extract geometry from images and compute interpretable physical metrics. The common thread is not shared implementation, but a recurring strategy: transform an accessible visual representation into a structured downstream model—whether a clean sketch, a near-surface transport proxy, or an energy-maintenance diagnostic.