ShadowDraw: From Any Object to Shadow-Drawing Compositional Art

1. Introduction

• Unlike prior computational shadow-art approaches that reproduce a predefined target, ShadowDraw aims for serendipitous compositions where shadows and drawings form a cohesive whole.The work takes inspiration from Vincent Bal's everyday-object shadow drawings.Source: §1 · p. 2
• The task must infer a subject without knowing it in advance, while rendered shadow inputs provide weak structural cues and shadow–drawing examples are scarce.These factors make high-quality, well-aligned generation challenging.Source: §1 · p. 2
• Shadow contours guide a text-conditioned drawing generator; differentiable scene search finds meaningful shadows, VLMs describe intended drawings, and automatic metrics filter for coherence and visual quality.The contour also supports scalable data construction and use of pretrained edge-conditioned generators.Source: §1 · p. 2
• The framework produces compelling compositions across 3D datasets, real-world scans, and generated assets, and extends to multi-object scenes, animation, and physical setups.The physical setup requires an everyday object, a planar surface, and a spotlight.Source: §1 · p. 2

2. Related Work

• Computational visual art often uses inverse design to match a target visual effect by optimizing scene parameters, whereas ShadowDraw also estimates the target subject and joins shadows with line drawings.Traditional shadow art treats objects and their cast shadows as the sole medium.Source: §2 · p. 3, Figure 2 · p. 3
• Prior line-drawing research spans geometry-based extraction and image- or text-guided generation, but generally treats drawing as an isolated modality rather than combining it with cast shadows.Source: §2 · p. 3
• Earlier shadow-art systems matched desired silhouettes by deforming or optimizing geometry [18, 32] [34] [36], but could yield impractical shapes; later work added multilayer occluders and colored shadows.Source: §2 · p. 3
• Shadow art has also used human bodies and hand gestures [44]; ShadowDraw instead pairs physical shadows with generative models that complete even a bare suggestion into a composition.Source: §2 · p. 3

3. Generative Shadow-Drawing Art

• A 3D object, partial line drawing, and scene parameters are arranged so the cast shadow completes a recognizable image.The setup uses a ground-plane canvas, spotlight, black drawing, and gray shadow silhouette; the study restricts scenes to a single light and omits surface textures.Source: §3 · p. 3

3.1. From Shadow Contour to Shadow-Drawing Art

• Replacing raw shadows with clean binary contours strengthens drawing alignment and enables edge-conditioned generation and scalable training-data synthesis.The training set includes 10K synthesized line drawings, with closed contours extracted as shadow-contour conditions.Source: §3.1 · p. 4
• At inference, the rendered shadow contour and text prompt condition drawing generation, while outpainting with an object mask prevents strokes from overlapping the object.The input contour is erased from the generated drawing before the object is reinserted and the shadow-drawing composition rendered.Source: §3.1 · p. 4

3.2. Scene Configuration Selection

• Scene selection optimizes light azimuth and elevation plus object position and rotation, using shadow fractal dimension as a differentiable measure of contour complexity.The search starts from 48 configurations and restricts updates to each initialization’s local neighborhood to keep configurations physically plausible and reproducible.Source: §3.2 · pp. 4–5
• A vision–language model proposes detailed, shadow-specific drawing prompts and allows users to control the intended subject.Prompts describe how the shadow contour can serve as a structural element in the drawing.Source: §3.2 · p. 5

3.3. Evaluation and Ranking

• Candidate compositions are filtered for shadow–drawing coherence and meaningful shadow contribution, then ranked by semantic and quality metrics for top-K selection.A VQA check discards strokes that do not outline the described component; comparisons of full and contour-removed drawings assess whether the shadow improves quality.Source: §3.3 · p. 5

4.1. Experimental Setup

• The evaluation covers 200 object models and compares ShadowDraw with two Gemini baselines, alongside ablations of conditioning, training data, and scene-parameter optimization.Metrics include CLIP score, ImageReward, Human Preference Score, and concealment; two user studies assess preference and satisfactory outputs.Source: §4.1 · pp. 5–6

4.2. Results and Analyses

• 70.4% of user-study cases favored ShadowDraw, and 96.8% of top-four result sets included at least one user-rated satisfactory image; Gemini baselines had low concealment.The results also report that participants marked 20.1% of comparisons as “neither preferable.”Source: §4.2 · p. 6, Figure 5 · p. 6, Table 1 · p. 6
• Shadow-contour conditioning improves generation quality and concealment, synthetic training data boosts all metrics, and scene-parameter optimization delivers the strongest overall performance.Qualitative ablations also show stronger shadow–drawing coherence and visual quality for the full method.Source: §4.2 · p. 6, Table 2 · p. 8, Figure 7 · p. 8

4.3. Applications

• The framework generates diverse works from one object, stacks multiple objects into composite scenes, and handles animated objects by combining five key-frame shadow contours.For animation, distinct colors mark the frames, and strokes are restricted to the intersection of shadow regions and their neighborhoods.Source: §4.3 · pp. 6, 8, Figure 6 · p. 7
• Physical deployment uses everyday objects and a spotlight; a phone flashlight and household items suffice for real-world compositions.The paper also describes physical prototypes of the letters C, V, P, and R.Source: §4.3 · pp. 6, 8

• The framework combines scene-parameter optimization, shadow-contour-guided line drawing, and automatic evaluation and ranking to produce unified shadow–drawing art.Source: §5 · p. 8
• Experiments report results across diverse 3D assets, with extensions to multi-object scenes, animations, and real-world setups.Source: §5 · p. 8
• Some objects produce uninformative shadows, diffusion inference adds runtime overhead, and automated ranking sometimes still requires human judgment.Source: §5 · p. 8

• The supplement provides additional implementation details and extended qualitative results.Source: §1 · p. 11

2.1. Training Data Construction

• The training set pairs 10K FLUX-generated line drawings with closed contours extracted and greedily merged from those drawings.The process begins with 100 GPT-4o drawings, then fine-tunes FLUX-1-dev LoRA on filtered drawings before synthesizing the additional 10K.Source: §2.1 · p. 11, Figure 8 · p. 11

2.2. Visual Prompt Proposal

• The visual prompt proposal uses a VLM to describe a drawing in which the shadow contour serves as a structural element, optionally guided by a user-specified subject.A second VLM checks whether the highlighted contour outlines the described component; candidates judged incoherent are discarded.Source: §2.2 · p. 11

2.3. Baseline Execution Details

• The baseline comparison prompts Gemini Flash 2.5 Image with either an object–shadow composite or a shadow contour, then renders the result with the object’s cast shadow.Source: §2.3 · p. 12

2.4. Discussion on the Evaluation Algorithm

• The ranking process filters incoherent cases and favors compositions where the shadow meaningfully contributes to the drawing.It combines VQA-based coherence verification with an assessment comparing the complete drawing against a contour-removed version.Source: §2.4 · p. 12, Figure 9 · p. 13
• 63.5% agreement, 11.0% disagreement, and 25.5% no preference in study one; top-three agreement was 39.8%, versus 44.5% between annotator sets.The studies collected 2,000 preference pairs from 10 annotators in the first study and two labels per top-three pair across 200 objects in the second.Source: §2.4 · p. 12

2.5. Subject-Specified Generation

• Prompt editing enables user-specified subjects, but some object geometries cannot provide a shadow stroke that integrates meaningfully with the requested concept.Source: §2.5 · p. 12, Figure 10 · p. 14

2.6. Animated Shadow-drawing Art

• For animation, five keyframe shadow contours are overlaid by frame, and a binary mask confines strokes to stable shadow regions; resulting drawings remain coherent across dynamic shadows.Pixels nearer to the dynamic than the static shadow region are masked out during synthesis.Source: §2.6 · pp. 12–13, Figure 11 · p. 14

2.7. Runtime Analysis

• The full pipeline takes about 35 minutes per object on one A6000 GPU; reducing sampling from 30 to 10 steps lowers this to around 15 minutes with minimal quality degradation.Parallel execution on standard 8-GPU nodes can reduce latency to under five minutes.Source: §2.7 · p. 15

3. Limitations

• Some object shapes produce shadows too uninteresting or ambiguous to interpret, regardless of lighting or pose, limiting the compositions the method can produce.Source: §3 · p. 15
• Joint search and generation over scene parameters creates noticeable runtime overhead, and a single-object result still takes a relatively long time.Source: §3 · p. 15
• The ranking algorithm is imperfect, so users may still need to inspect multiple outputs to find the most compelling result.Source: §3 · p. 15