Minecraft AI Suite
Collection
Minecraft datasets for vision AI — from classification to generation. Multiple resolutions, multiple tasks. More coming. • 4 items • Updated • 1
🧱 Latent Flow Matching Minecraft 256×256
Unconditional generative model trained from scratch with Flow Matching on latents from Stable Diffusion's VAE.
Generates Minecraft-style landscape screenshots.
🏗️ Architecture
| Component | Details |
|---|---|
| UNet | UNet2DModel (~17M params), 4 levels, attention at level 2 |
| VAE | for latent encode/decode |
| Latent | 32×32×4 (256×256 image, downsample 8) |
| Training | Flow Matching (x_t = (1-t)·x_0 + t·noise, predicts velocity noise - x_0) |
| Sampler | Euler ODE, 30 steps |
🚀 Usage
See generate.py.
📊 Data
- Custom screenshots taken in Minecraft with Photon Shaders by SixthSurge.
- ~10,500 images, 256×256 random crops from Full HD.
- Augmentations: random crop, horizontal flip.
📈 Training progress
Epoch 20
Epoch 40
Epoch 60
Epoch 80
Epoch 100
📜 License
CC BY-NC-SA 4.0. Commercial use prohibited.
🙏 Credits
- Minecraft is a trademark of Mojang Synergies AB and Microsoft.
- Photon Shaders © Benjamin Stott (SixthSurge), used under its license permitting commercial screenshots.
See
PHOTON_SHADERS_LICENSE.txt. - VAE:
stabilityai/sd-vae-ft-mse(MIT).
- Downloads last month
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for Nininkkka/Latent-Flow-Matching-Minecraft-256x256
Base model
stabilityai/sd-vae-ft-mseCollection including Nininkkka/Latent-Flow-Matching-Minecraft-256x256
Evaluation results
- FID on Minecraft Screenshots (Photon Shaders)self-reported108.990
- Inception Score on Minecraft Screenshots (Photon Shaders)self-reported5.420
- KID on Minecraft Screenshots (Photon Shaders)self-reported0.069