Research map / Image generation / Autoregressive and latent generation
LookingGlass: Generative Anamorphoses via Laplacian Pyramid Warping
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.
Generates text-conditioned anamorphic illusions with latent rectified-flow models (Stable Diffusion 3.5) by synchronizing predicted clean images across mirror/lens views and warping them with a Laplacian pyramid, achieving better quality than Visual Anagrams-style baselines.
Problem
Creating ambiguous anamorphic images that show one meaning when viewed directly and a different hidden image through a mirror or lens, which prior methods could only do in pixel space with orthogonal transforms and at lower quality.
Approach
Extends synchronized multi-view denoising to latent rectified-flow models by averaging predicted clean images in pixel space via VAE round-trips with residual correction, and introduces Laplacian Pyramid Warping, an LOD-aware frequency-preserving warp that blends views at each frequency band.
Key results
- On 135° rotation, LookingGlass reaches FID 129.74 vs 166.03 for SyncTweedies and 293.00 for Visual Anagrams (Geng et al.).
- On cylindrical mirror anamorphosis, LookingGlass achieves FID 130.27 and alignment A 0.272, versus 0.171 for Geng et al.
- A 27-participant user study consistently preferred LookingGlass results over prior methods; each image pair takes about 80 seconds on an RTX 4090 with 30 steps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the predicted clean image estimate instead of predicted noise (ablation) | Visual Anagrams with Stable Diffusion 3 · Vertical Flip | FID: 335.24 -> 171.61 | Helped |
| add VAE encoding/decoding for image-space transformations (ablation) | Visual Anagrams with Stable Diffusion 3 · Vertical Flip | FID: 171.61 -> 160.17 | Helped |
| add Laplacian Pyramid Warping (ablation) | LookingGlass with Stable Diffusion 3 · Vertical Flip | FID: 160.17 -> 153.27 | Helped |
| add time traveling during intermediate denoising steps (ablation) | LookingGlass with Stable Diffusion 3 · Vertical Flip | FID: 153.27 -> 150.01 | Helped |
| add time traveling during intermediate denoising steps (ablation) | LookingGlass with Stable Diffusion 3 · Vertical Flip | CLIP alignment score A: 0.300 -> 0.295 | Hurt |
| use LookingGlass instead of SyncTweedies | 135° Rotation · 135° Rotation | FID: 166.03 -> 129.74 | Helped |
| use LookingGlass instead of SyncTweedies | Cylindrical Mirror · Cylindrical Mirror | CLIP alignment score A: +0.031 | Helped |
| use LookingGlass instead of SyncTweedies | Vertical Flip · Vertical Flip | CLIP alignment score A: 0.302 -> 0.297 | Hurt |
Surprising findings
- Naive VAE-based latent synchronization (Tancik) fails badly on rotations, with FID 439 at 135°, showing latent models need special handling.
Limitations
- Not all prompt pairs yield high-quality illusions; prompt selection remains a challenge.
- Computationally more expensive than pixel-space Visual Anagrams due to VAE encode/decode steps.
- Evaluation of CLIP alignment and concealment scores uses only 50 prompt pairs.
Experimental setup
Metrics: CLIP alignment score concealment score FID KID user preference ranking
Baselines: Visual Anagrams Diffusion Illusions SyncTweedies Tancik SD 3.5
Builds on: Visual Anagrams SyncTweedies Diffusion Illusions Generative Powers of Ten FreeDoM
Closest papers by meaning
- Visual Anagrams: Generating Multi-View Optical Illusions with Diffusion Models (CVPR 2024 · similarity 0.89)
- PTDiffusion: Free Lunch for Generating Optical Illusion Hidden Pictures with Phase-Transferred Diffusion Model (CVPR 2025 · similarity 0.84)
- Pygmalion Effect in Vision: Image-to-Clay Translation for Reflective Geometry Reconstruction (NeurIPS 2026 · similarity 0.84)
- MirrorVerse: Pushing Diffusion Models to Realistically Reflect the World (CVPR 2025 · similarity 0.84)
- Generative Perception of Shape and Material from Differential Motion (NeurIPS 2025 · similarity 0.84)
- Multistable Shape from Shading Emerges from Patch Diffusion (NeurIPS 2024 · similarity 0.83)
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.