Research map / Image generation / Autoregressive and latent generation
Terrain Diffusion Network: Climatic-Aware Terrain Generation with Geological Sketch Guidance
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 Terrain Diffusion Network (TDN), a sketch-guided latent diffusion model with three separately trained synthesisers (structural, intermediate, fine-grained) for controllable, climate-aware terrain generation, evaluated on a new NASA-topology-derived dataset.
Problem
Sketch-guided terrain generation struggles to combine flexible user control with fine-grained, realistic (climate/erosion-shaped) terrain detail and diversity; GAN-based methods trade fidelity against adherence to user conditions, and general diffusion models are poorly suited to terrain guidance.
Approach
Replace a monolithic denoiser with a coarse-to-fine multi-level scheme of three separately trained U-Net synthesisers operating in terrain and sketch latent spaces from pre-trained autoencoders, with sketch latents concatenated directly to the noisy latent instead of cross-attention conditioning.
Key results
- TDN achieves FID 0.44023 and MSE 0.00590, versus 4.66 FID / 0.0391 MSE for the best baseline (GAN).
- Removing the multi-level synthesisers raises FID to 0.50707 and MSE to 0.00707; removing the intermediate synthesiser raises FID to 1.88794.
- Removing the terrain and sketch autoencoders gives FID 1.94008 and MSE 0.03118.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TDN instead of GAN | Terrain generation · NASA terrain test set | FID: 4.65986 -> 0.44023 | Helped |
| Use TDN instead of GAN | Terrain generation · NASA terrain test set | MSE: 0.03910 -> 0.00590 | Helped |
| Add multi-level synthesisers rather than using a single denoiser (ablation) | TDN · NASA terrain test set | FID: 0.50707 -> 0.44023 | Helped |
| Add the intermediate-level synthesizer (ablation) | TDN · NASA terrain test set | FID: 1.88794 -> 0.44023 | Helped |
| Add the terrain and sketch autoencoders (ablation) | TDN · NASA terrain test set | FID: 1.94008 -> 0.44023 | Helped |
| Use direct sketch guidance instead of cross-attention (ablation) | TDN · NASA terrain test set | FID: 1.89567 -> 0.44023 | Helped |
Surprising findings
- A single synthesiser sometimes outperformed the two-synthesiser variant, suggesting the number of synthesisers alone does not determine performance.
- Stable Diffusion, ControlNet and GLIGEN were worse than the GAN baseline on FID, which the authors attribute to lack of terrain-specific design.
Limitations
- Generation quality depends on training data scale and diversity; the model can only denoise samples within the training distribution.
- Training the autoencoders independently and then the diffusion model makes overall training computationally expensive; inference is slow (about 11 s).
- Control is limited to a fixed set of four sketch types (rivers, ridges, basins, peaks) and no text or flexible guidance.
Experimental setup
Datasets: NASA Topology Images (custom terrain dataset)
Benchmarks: custom NASA Topology test split (2,611 samples)
Metrics: FID MSE
Baselines: GAN VQGAN Stable Diffusion ControlNet GLIGEN
Builds on: denoising diffusion probabilistic models classifier-free guidance Stable Diffusion e-Diff
Closest papers by meaning
- Denoising Diffusion via Image-Based Rendering (ICLR 2024 · similarity 0.84)
- SatSynth: Augmenting Image-Mask Pairs through Diffusion Models for Aerial Semantic Segmentation (CVPR 2024 · similarity 0.84)
- Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.84)
- De-coupled NeuroGF for Shortest Path Distance Approximations on Large Terrain Graphs (ICML 2025 · similarity 0.83)
- Understanding the Generalization of Pretrained Diffusion Models on Out-of-Distribution Data (AAAI 2024 · similarity 0.83)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · 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.