Research map / Image generation / Autoregressive and latent generation
Tiled Diffusion
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.
Tiled Diffusion extends pretrained latent diffusion models (mainly Stable Diffusion 1.5/2.0, plus SDXL) with training-free latent-space tiling and similarity constraints to generate multiple seamlessly connected images for self-, one-to-one, and many-to-many tiling.
Problem
Existing automated tiling methods mostly handle single-image self-tiling or texture synthesis and cannot generate multiple interconnected tiles with arbitrary side-to-side connections across diverse image domains.
Approach
At every diffusion step, copy latent context from the connected sides (with rotation and round-robin selection) into padding regions, plus a similarity constraint that copies a narrow edge window so many-to-many connections match, letting all tiles be generated jointly and share information.
Key results
- In self-tiling, Tiled Diffusion achieves TS 0.03 with FID 47.9, versus STI at TS 0.03 with FID 59.2 and AT at FID 49.2.
- In one-to-one tiling, Tiled Diffusion reaches TS 0.04 and FID 48.1, while STI has TS 0.14 and FID 77.1.
- Only Tiled Diffusion supports many-to-many tiling, reaching TS 0.03 and FID 50.5 in that setting; ablation without the tiling constraint yields TS 0.28-0.31 (untiled).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the tiling constraint (ablation) | Tiled Diffusion · Self-tiling | TS: 0.28 -> 0.03 | Helped |
| Add the tiling constraint (ablation) | Tiled Diffusion · One-to-one tiling | TS: 0.31 -> 0.04 | Helped |
| Add the tiling constraint (ablation) | Tiled Diffusion · Many-to-many tiling | TS: 0.30 -> 0.03 | Helped |
| Add the similarity constraint (ablation) | Tiled Diffusion · Self-tiling | TS: 0.03 -> 0.03 | No clear effect |
| Add the similarity constraint (ablation) | Tiled Diffusion · One-to-one tiling | TS: 0.04 -> 0.04 | No clear effect |
| Add the similarity constraint (ablation) | Tiled Diffusion · Many-to-many tiling | TS: 0.12 -> 0.03 | Helped |
| Replace STI with Tiled Diffusion | Stable Diffusion self-tiling · Self-tiling | FID: 59.2 -> 47.9 | Helped |
| Replace STI with Tiled Diffusion | Stable Diffusion one-to-one tiling · One-to-one tiling | TS: 0.14 -> 0.04 | Helped |
| Replace STI with Tiled Diffusion | Stable Diffusion one-to-one tiling · One-to-one tiling | FID: 77.1 -> 48.1 | Helped |
| Replace AT with Tiled Diffusion | Stable Diffusion self-tiling · Self-tiling | FID: 49.2 -> 47.9 | Helped |
| Replace AT with Tiled Diffusion | Stable Diffusion self-tiling · Self-tiling | TS: 0.03 -> 0.03 | No clear effect |
Surprising findings
- Removing the tiling constraint yields images with TS similar to non-tiled vanilla generation, confirming the constraint is what produces tileability.
Limitations
- Cross-axis (X-to-Y) connections show small artifacts because latent rotations do not perfectly match pixel-space rotations.
- Evaluation is limited to a single LAION caption sample and pseudo ground truth built from Stable Diffusion outputs.
Experimental setup
Datasets: LAION 400M captions
Benchmarks: LAION 400M captions (1,000 sampled)
Metrics: Tiling Score (TS) CLIP-Score LPIPS FID
Baselines: Seamless Tile Inpainting (STI) Asymmetric Tiling (AT) vanilla text-to-image generation (VT2I)
Builds on: Stable Diffusion Differential Diffusion Asymmetric Tiling
Closest papers by meaning
- Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.86)
- StochSync: Stochastic Diffusion Synchronization for Image Generation in Arbitrary Spaces (ICLR 2025 · similarity 0.86)
- DreamCraft3D: Hierarchical 3D Generation with Bootstrapped Diffusion Prior (ICLR 2024 · similarity 0.85)
- Kiss3DGen: Repurposing Image Diffusion Models for 3D Asset Generation (CVPR 2025 · similarity 0.85)
- MatFuse: Controllable Material Generation with Diffusion Models (CVPR 2024 · similarity 0.85)
- GenesisTex: Adapting Image Denoising Diffusion to Texture Space (CVPR 2024 · 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.