Terrain Diffusion MC Linux AMD Fork

Terrain Diffusion MC for AMD GPUs on Linux through ROCm/MIGraphX.

Upstream Credit

All core Terrain Diffusion Minecraft integration work comes from the original project and its contributors:

Original Minecraft mod: xandergos/terrain-diffusion-mc
Original Terrain Diffusion models/project: xandergos/terrain-diffusion

This fork exists to document and package an AMD/Linux-oriented bring-up. Upstream authors deserve the credit for the original mod, model integration, and most of the codebase.

Fork Scope

This fork adds and documents:

Linux AMD GPU provider selection via migraphx
explicit ONNX Runtime native library override support
a dedicated linux-amd build variant
notes for a tested ROCm/MIGraphX setup

It does not replace upstream. If you want the original project and its normal release stream, use the upstream repository.

Verified Setup

OS: Kubuntu 26.04 LTS
Kernel: Linux 7.0.0-27-generic
CPU: AMD Ryzen 7 9700X (8 cores / 16 threads)
GPU: AMD Radeon RX 7900 XTX (gfx1100)
System memory: 64 GB
ROCm: 7.2.0
Java: OpenJDK 21.0.11
Minecraft: 1.21.1
NeoForge: 21.1.238
ONNX Runtime: custom 1.23.2 build with MIGraphX Execution Provider

The Minecraft runtime log confirms Terrain Diffusion (MIGraphX) during 256x256 region generation. The project build targets NeoForge 21.1.229; the tested NeoForge 21.1.238 runtime is compatible with that target.

Build Matrix

Build	Supports	Setup required
Linux AMD	Linux + AMD GPU + ROCm/MIGraphX	Custom MIGraphX-enabled ONNX Runtime natives
Windows	Windows with DirectML-capable GPU	None beyond the bundled DirectML runtime
CUDA	Linux/Windows with NVIDIA GPU	CUDA + cuDNN
CPU	Any machine	None, but very slow

For AMD on Linux, use the linux-amd jar from this fork's releases, not the CPU jar and not the upstream release artifacts.

Requirements

Minecraft with NeoForge
Java 21
For the Linux AMD path:
- an AMD GPU supported by ROCm/MIGraphX
- a working ROCm userspace install
- custom ONNX Runtime native libraries built with MIGraphX enabled

Memory requirements vary with world scale and chunk generation pressure. On the tested machine:

GPU VRAM used by inference: roughly 1.5-2 GB
Minecraft RAM allocation used for testing: 12 GB

Quick Start

Install the linux-amd jar in your Minecraft mods/ folder.
Launch once so the mod can create its config and download models.
Edit config/terrain-diffusion-mc.properties.
Set:

inference.device=gpu
inference.provider=migraphx
inference.native_path=/absolute/path/to/onnxruntime-migraphx-native-directory

That directory contained:

libonnxruntime.so
libonnxruntime4j_jni.so
libonnxruntime_providers_migraphx.so
libonnxruntime_providers_shared.so

Configuration

The mod config file is config/terrain-diffusion-mc.properties.

Relevant inference settings in this fork:

# "cpu", "gpu", or "auto"
inference.device=gpu

# "auto", "migraphx", "rocm", "cuda", or "directml"
inference.provider=migraphx

# Optional path to custom ONNX Runtime native libraries
inference.native_path=

# Offload inactive models between stages
inference.offload_models=true

# Validate existing model files
validate_model=true

# Explorer web UI port
explorer.port=19801

# Chunk generation region size in blocks
tile_size=256

World Scale Notes

Terrain Diffusion worlds expose a World Scale setting from 1 to 6.

2 is the default and the most balanced setting in this codebase.
Higher values are not "more detail". They generate at native model resolution and then upsample more aggressively.
Above 4, very broad smoothing can produce large ocean-heavy starts. That is expected from the current scaling logic, not specifically an AMD/Linux issue.

Dependencies

Primary runtime/build dependencies involved in this fork:

Minecraft 1.21.1
NeoForge 21.1.229
ONNX Runtime 1.23.2
ROCm/MIGraphX 7.2.x

The Linux AMD path in this fork relies on a custom ONNX Runtime Java/native build because the stock Maven Java artifacts do not ship as a ready-made MIGraphX package for this workflow.

Building From Source

Standard upstream variants still work:

./gradlew build -PuseDml=true
./gradlew build -PuseCuda=true
./gradlew build -PuseCpu=true

Linux AMD build for this fork:

./gradlew build -PuseMigraphx=true -PonnxRuntimeMigraphxVersion=1.23.2-migraphxfix

This expects a matching patched ONNX Runtime Java artifact to be available through mavenLocal().

Linux AMD Notes

Japanese installation guide: docs/INSTALL_JA.md.

Detailed notes for the AMD/Linux port are in docs/amd-linux-migraphx.md.

Common Issues

No configured GPU provider is available

This usually means one of these is wrong:

the jar does not include the MIGraphX-capable Java bindings
inference.native_path does not point at the custom ORT native library directory
the native library directory is missing provider .so files
ROCm/MIGraphX is not installed or not loadable on the machine

Massive ocean starts at high world scale

This is an effect of the current upsampling-based scale system. Reduce World Scale back to 2 or 3 if you want less smoothing and more compact terrain features.

Out-of-memory during world generation or model load

Increase Minecraft RAM allocation and keep world scale moderate. The model files themselves are large, and higher scale values change chunk-generation behavior in ways that can still stress memory and CPU.

Terrain Diffusion Forked(Unofficial)