Research map / Image generation / Text-to-image diffusion
Visual Anagrams: Generating Multi-View Optical Illusions with Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a zero-shot method that generates multi-view optical illusions (flips, rotations, permutations, inversions) by averaging noise estimates across transformed views during reverse diffusion of an off-the-shelf text-to-image model.
Problem
Generating images whose appearance changes under a transformation (flip, rotation, permutation) such that each view matches a different text prompt, without an explicit model of human perception.
Approach
Transform the noisy image under each view, estimate noise with a pretrained diffusion model conditioned on each view's prompt, map estimates back with the inverse view, and average them for a single denoising step. A theory shows this is valid exactly when views are orthogonal transformations (which preserve Gaussian noise statistics), and pixel-space diffusion avoids latent-patch artifacts.
Key results
- On the hand-made prompt dataset with vertical flips, the method reaches alignment score 0.275 and concealment score 0.574, versus 0.233/0.501 for Burgert et al. (SDS) and 0.256/0.545 for Tancik.
- On CIFAR prompt pairs, alignment 0.287 and concealment 0.624 vs. 0.225/0.501 (Burgert) and 0.278/0.595 (Tancik).
- Mean reduction of noise estimates outperforms alternating reduction (A 0.255 vs 0.252; C 0.567 vs 0.560), and higher guidance scale (gamma 10) gave the best concealment among tested values.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed method instead of Tancik | two-view illusion generation · CIFAR | alignment score: 0.278 -> 0.287 | Helped |
| use the proposed method instead of Tancik | two-view illusion generation · CIFAR | concealment score: 0.595 -> 0.624 | Helped |
| use negative prompting (ablation) | Visual Anagrams · Ours | alignment score: 0.255 vs 0.24 | Hurt |
| use negative prompting (ablation) | Visual Anagrams · Ours | concealment score: 0.567 vs 0.576 | Helped |
| replace mean reduction with alternating reduction (ablation) | Visual Anagrams · Ours | alignment score: 0.255 vs 0.252 | Hurt |
| replace mean reduction with alternating reduction (ablation) | Visual Anagrams · Ours | concealment score: 0.567 vs 0.560 | Hurt |
| set guidance scale to γ = 3.0 rather than γ = 7.0 (ablation) | Visual Anagrams · Ours | alignment score: 0.239 vs 0.255 | Hurt |
| set guidance scale to γ = 3.0 rather than γ = 7.0 (ablation) | Visual Anagrams · Ours | concealment score: 0.537 vs 0.567 | Hurt |
| set guidance scale to γ = 10.0 rather than γ = 7.0 (ablation) | Visual Anagrams · Ours | alignment score: 0.259 vs 0.255 | Helped |
| set guidance scale to γ = 10.0 rather than γ = 7.0 (ablation) | Visual Anagrams · Ours | concealment score: 0.576 vs 0.567 | Helped |
Surprising findings
- Pixel-space DeepFloyd IF avoids artifacts that latent diffusion (Stable Diffusion) produces under rotations and flips, where straight lines become thatched.
- Even randomly sampled orthogonal transformations (16384-dim) yield reasonable-looking, if incomprehensible, illusions, confirming the theory.
Limitations
- Views must be orthogonal (volume-preserving); non-volume-preserving deformations, homographies, stretches and color-constancy illusions are unsupported.
- Does not consistently produce perfect illusions; some outputs show independent synthesis or noise-shift artifacts.
- Bilinear rotation introduces noise correlations causing divergent samples.
Experimental setup
Datasets: CIFAR-10 class prompt pairs hand-compiled prompt pairs (50 pairs)
Benchmarks: CIFAR prompt pairs (45) hand-compiled prompt pairs (50)
Metrics: alignment score A (min diagonal CLIP score) concealment score C (CLIP softmax trace) quantiles A0.9, A0.95, C0.9, C0.95
Baselines: Burgert et al. score distillation sampling Tancik rotation illusion method
Builds on: DDPM DDIM classifier-free guidance compositional diffusion (energy-based composition) Tancik multi-view rotation illusion
Closest papers by meaning
- LookingGlass: Generative Anamorphoses via Laplacian Pyramid Warping (CVPR 2025 · similarity 0.89)
- PTDiffusion: Free Lunch for Generating Optical Illusion Hidden Pictures with Phase-Transferred Diffusion Model (CVPR 2025 · similarity 0.88)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.86)
- ViVid-1-to-3: Novel View Synthesis with Video Diffusion Models (CVPR 2024 · similarity 0.86)
- The Art of Deception: Color Visual Illusions and Diffusion Models (CVPR 2025 · similarity 0.86)
- MirrorVerse: Pushing Diffusion Models to Realistically Reflect the World (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.