Talus
A lightweight diffusion model with 23 million parameters, designed for procedural generation of game terrain heightmaps. It supports in-browser inference via WebGPU, employs a U-Net architecture, and uses DDIM sampling for efficient generation.
Timeline (last 90 days)
Talus is a diffusion model with only 23 million parameters that can generate game terrain heightmaps in real time in the browser via WebGPU.
Talus is an open-source project published on Reddit by developer osfv.
Talus generates 64×64 heightmaps corresponding to terrain covering approximately 4 kilometers in real-world scale with a maximum elevation of 1,200 meters.
Talus's generation process can be controlled by terrain type and any subset of five measurable attributes: mean elevation, relief, mean slope, water fraction, and spectral slope.
Each attribute is paired with a learnable unknown embedding vector and is independently randomly dropped during training, allowing the model to work at inference time regardless of which attributes are specified or left blank.
The Talus model was trained from scratch on a single RTX 5060 (8 GB VRAM), accumulating approximately 4.5 hours of training across multiple versions.
The training data consists of 45,000 maps produced by the author using a procedural generator, with a generation pipeline incorporating terrain simulation algorithms such as fBm/ridged noise, river erosion, slope diffusion, and thermal erosion.
The Talus model architecture uses a pixel-space U-Net, v-prediction objective, cosine noise schedule, 50-step DDIM sampling with quadratic spacing, and classifier-free guidance with a strength of 2.0.
By generating relative-height terrain shapes normalized by relief and placing them at the target elevation via the sampler, the plain distance ratio was reduced from 3.98 to 1.23.
Talus's evaluation method divides all distance metrics by the distance between two disjoint sets of real maps as a noise baseline, where 1.0 means the generated results are indistinguishable from real maps.
5 more timeline events
All Facts (15)
Small diffusion models are generally poor at high-frequency details, and Talus struggles to reproduce ridgelines and the finest spectral bands, with mountains being over-smoothed and plains appearing too grainy.
50%UnverifiedTalus is an open-source project published on Reddit by developer osfv.
50%UnverifiedTalus generates 64×64 heightmaps corresponding to terrain covering approximately 4 kilometers in real-world scale with a maximum elevation of 1,200 meters.
50%UnverifiedTalus's generation process can be controlled by terrain type and any subset of five measurable attributes: mean elevation, relief, mean slope, water fraction, and spectral slope.
50%UnverifiedEach attribute is paired with a learnable unknown embedding vector and is independently randomly dropped during training, allowing the model to work at inference time regardless of which attributes are specified or left blank.
50%UnverifiedThe Talus model was trained from scratch on a single RTX 5060 (8 GB VRAM), accumulating approximately 4.5 hours of training across multiple versions.
50%UnverifiedThe training data consists of 45,000 maps produced by the author using a procedural generator, with a generation pipeline incorporating terrain simulation algorithms such as fBm/ridged noise, river erosion, slope diffusion, and thermal erosion.
50%UnverifiedThe Talus model architecture uses a pixel-space U-Net, v-prediction objective, cosine noise schedule, 50-step DDIM sampling with quadratic spacing, and classifier-free guidance with a strength of 2.0.
50%UnverifiedBy generating relative-height terrain shapes normalized by relief and placing them at the target elevation via the sampler, the plain distance ratio was reduced from 3.98 to 1.23.
50%UnverifiedTalus's evaluation method divides all distance metrics by the distance between two disjoint sets of real maps as a noise baseline, where 1.0 means the generated results are indistinguishable from real maps.
50%UnverifiedThe current Talus model's performance on the test set is: metric W1 distance at 1.51× the baseline, spectral at 9.1×, and slope at 1.65×.
50%UnverifiedTalus is exported in ONNX format with weights stored in fp16 and converted back to fp32 on load, runs on WebGPU via ONNX Runtime Web, generates one map in approximately 3 seconds on an RTX 5060, and provides a CPU fallback.
50%UnverifiedThe author reimplemented the sampler in JavaScript and verified that its output differs from the PyTorch version by less than 0.6 meters on reference samples.
50%UnverifiedTalus is a diffusion model with only 23 million parameters that can generate game terrain heightmaps in real time in the browser via WebGPU.
50%UnverifiedThe Talus project has an online demo available at talus.tersa.tech, and its code, weights, and scorecard are open-sourced on GitHub under the Apache-2.0 license.
50%