VS Terrain Diffusion

<p><span style="color: #e03e2d;"><strong>This mod is in early development, and is provided here for the community to play around with and explore. World breaking changes should be expected. Only use this mod with a new world.</strong></span></p> <p>&nbsp;</p> <p><strong>I also recommend checking out this implementation of Terrain Diffusion by Alyis: <a href="https://mods.vintagestory.at/show/mod/61384">Terra Loom</a></strong></p> <p>&nbsp;</p> <p><strong>Description</strong></p> <p>Generates Vintage Story worlds with<a href="https://github.com/xandergos/terrain-diffusion"> Terrain Diffusion;</a><br />a neural heightmap model trained on real Earth topography.</p> <p>&nbsp;</p> <p><strong>Recent Updates</strong></p> <ul> <li>Many optimizations and support for OpenVINO thanks to <strong><a href="https://github.com/ThePaul-T">Paul T</a>. (v 0.8.1)</strong></li> <li>Fixed translocators so they find a destination in a reasonable amount of time.<strong> (v 0.9.1)</strong></li> <li>Added native, subway tunnel free, support for rivers via model conditioning.<strong> (v 0.9.1)</strong></li> <li>Changed model inferrence to be preemptable, so moving fast or teleporting will start generating terrain near the player fairly quickly.<strong> (v 0.9.1)</strong></li> <li>Descreased the default value for RiverBasinDepth to 0.3 to prevent river basins from flattening terrain too much.<strong> (v 0.9.2)</strong></li> <li>Improved reliability of hardware capability detection &amp; inference backend selection<strong> (v 0.9.3)</strong></li> <li>Windows will now (<strong>hopefully</strong>) work with TensorRT-RTX without crashing<strong> (v 0.9.3)</strong></li> <li>Incompatible configurations can no longer be selected in the configlib menu<strong> (v 0.9.3)</strong></li> <li>Fixed terrain flattening near sea level.<strong> (v 0.10.0)</strong></li> </ul> <p>&nbsp;</p> <p><strong>Known Issues</strong></p> <ul> <li>The very large biomes can make gameplay difficult, some easy tuning knobs are needed here.</li> <li>This mod is completely untested on macOS (I dont have a system to do so), and coreml support only works in theory.</li> <li><span style="text-decoration: line-through;">The int8 decoder model is being auto selected on windows if a gpu inferrence device is not detected. The Int8 decoder should require explicit selection. (<strong>High priority</strong>)</span></li> <li><span style="text-decoration: line-through;">River conditioning is too strong in areas with low terrain, causing wide swaths of perfectly flat terrain.</span></li> </ul> <p>&nbsp;</p> <p><strong>Requirements</strong></p> <ul> <li>~2.3 GB of disk for the model files, downloaded once on first launch</li> <li>~3 GB of extra VRAM while the models are resident</li> </ul> <p>&nbsp;</p> <table style="border-collapse: collapse; width: 100%; height: 80.8px;" border="1"> <tbody> <tr style="height: 20.2px;"> <td style="width: 33.3052%; height: 20.2px;">Platform</td> <td style="width: 33.3052%; height: 20.2px;">Device used automatically</td> <td style="width: 33.3052%; height: 20.2px;"> <p>Notes</p> </td> </tr> <tr style="height: 20.2px;"> <td style="width: 33.3052%; height: 20.2px;">Linux</td> <td style="width: 33.3052%; height: 20.2px;">CUDA</td> <td style="width: 33.3052%; height: 20.2px;">Needs CUDA runtime 12 or 13 and cuDNN9</td> </tr> <tr style="height: 20.2px;"> <td style="width: 33.3052%; height: 20.2px;">Windows</td> <td style="width: 33.3052%; height: 20.2px;">DirectML</td> <td style="width: 33.3052%; height: 20.2px;">Any modern GPU; no extra install</td> </tr> <tr style="height: 20.2px;"> <td style="width: 33.3052%; height: 20.2px;">macOS</td> <td style="width: 33.3052%; height: 20.2px;">CoreML</td> <td style="width: 33.3052%; height: 20.2px;">No extra install</td> </tr> <tr> <td style="width: 33.3052%;">Anything else</td> <td style="width: 33.3052%;">CPU</td> <td style="width: 33.3052%;">Works, but slow</td> </tr> </tbody> </table> <p>&nbsp;</p> <p>The matching ONNX Runtime native library is downloaded on first use (a few MB for CPU/DirectML,<br />~300 MB for CUDA). Nothing has to be installed by hand.</p> <p>&nbsp;</p> <p>If you have an NVIDIA 3000 series or newer, you can try out TensorRT-RTX (<strong>Tested with Linux and Win11, Win10 may have issues atm.</strong>). It will speed things up significantly.</p> <p>To enable it, set <em><strong>InferenceDevice</strong></em> to <em><strong>tensorrt-rtx</strong></em> in <strong>vsterraindiffusion.json</strong>.</p> <p>&nbsp;</p> <p><strong>Recommended Mods</strong></p> <ul> <li><a href="https://mods.vintagestory.at/terrainslabs">Terrain Slabs</a></li> <li><a href="https://mods.vintagestory.at/show/mod/22371">Farseer</a></li> <li><a href="https://mods.vintagestory.at/rivers">Rivers</a> (Set "RiverSpawnChance" in rivers.json to 0.1 or even 0.05)</li> <li><a href="https://mods.vintagestory.at/distantvistas">Distant Vistas</a> (If you have an extremely powerful computer, this mod is incredible)</li> </ul> <p>&nbsp;</p> <p><strong>Recommended World Settings</strong></p> <table style="border-collapse: collapse; width: 100%; height: 84.8px;" border="1"> <tbody> <tr style="height: 24.2px;"> <td style="width: 49.9578%; height: 24.2px;">Landcover</td> <td style="width: 49.9578%; height: 24.2px;">40% to 60%</td> </tr> <tr style="height: 20.2px;"> <td style="width: 49.9578%; height: 20.2px;">Landcover scale</td> <td style="width: 49.9578%; height: 20.2px;">300%</td> </tr> <tr style="height: 20.2px;"> <td style="width: 49.9578%; height: 20.2px;">World Height</td> <td style="width: 49.9578%; height: 20.2px;">512 (Generally the higher, the better. You can try 768 and 1024 as well. Climates and terrain may be a bit odd atm.)</td> </tr> <tr style="height: 20.2px;"> <td style="width: 49.9578%; height: 20.2px;">Color Accurate Map</td> <td style="width: 49.9578%; height: 20.2px;">Enabled</td> </tr> </tbody> </table> <p>&nbsp;</p> <p>The model generates terrain and climates that best fit the selected landcover (la