Faïence

Text-to-texture diffusion model for Minecraft. Generates 16×16 RGBA textures from a text prompt — sharp, pixel-exact, ready to drop into a resource pack.

Faïence is the French art of hand-painting glazed ceramic tiles, one at a time.

Installation

Requires Python 3.10+ and PyTorch (CUDA, CPU or MPS).

pip install torch
pip install -r requirements.txt

Download the latest weights from the GitHub release https://github.com/LeChatOTapas/faience-ai/releases/latest/download/faience.pth and place the file next to generate.py.

Usage

python generate.py --prompt "diamond sword" --upscale 16
python generate.py --prompt "ruby ore" --tile
python generate.py --prompt "emerald hammer" --sprite
python generate.py --prompt "sapphire block" --tile --num 6

Output PNGs are written to generated/. Without --upscale you get the native 16×16 file; --upscale 16 additionally scales it to 256×256 (nearest-neighbor) for inspection.

Prompt structure

Prompts are in English (the model is conditioned on CLIP text embeddings). Two structure keywords were part of the training captions and act as reliable controls:

Flag	Appended text	Effect
--tile	, full tile	Full-coverage block texture (blocks, ores, terrain)
--sprite	, sprite	Item on a transparent background

--tile is strongly recommended for blocks and ores. The attributes can also be written manually, together with color words, which the model follows: "glowstone, yellow, full tile", "black and red sword, sprite".

Options

Option	Default	Description
--num N	1	Variants per prompt
--seed N	random	Reproducible sampling
--steps N	50	Denoising steps
--guidance G	2.0	Text guidance strength (1.0 = unguided, 3.0 = literal)
--cfg_tmax T	500	Apply guidance only below this noise level (see below)
--colors N	off	Quantize palette to N colors
--upscale K	1	Nearest-neighbor upscaling factor
--prompt_file F	—	Batch generation, one prompt per line

Model

Faïence is a conditional pixel-space DDPM (~72M parameters). There is no VAE and no upscaler: denoising operates directly on the 16×16×4 RGBA grid, so every generated pixel is an actual pixel rather than a decoded approximation.

• U-Net (diffusers UNet2DConditionModel, 16→8→4) with cross-attention to a frozen CLIP ViT-B/32 text encoder — CLIP embeddings are what make unseen compositions work ("uranium ore", "obsidian crown").
• v-prediction with zero-terminal-SNR (cosine schedule), required for generation from pure noise at this resolution.
• Ancestral DDIM sampling (eta 1.0): reinjected noise preserves the grain of stochastic textures (ores, wool) instead of averaging it away.
• Interval guidance (--cfg_tmax): classifier-free guidance is applied only below a noise-level threshold, where the unconditional estimate is reliable. This permits strong guidance without the artifacts CFG normally produces under zero-terminal-SNR.
• Training data: ~14,500 textures at 16×16 (vanilla, mods, resource packs), with automatically enriched captions (dominant colors, tile/sprite structure) applied with dropout so both plain and attributed prompts stay in-distribution.

Known limitations

• 16×16 output only, by design.
• English prompts only.
• Block prompts generally need --tile for full-tile coverage.
• A few concepts remain unreliable (bows, notably). Use --num 4 and pick.

License

The code is released under the MIT License (see LICENSE). The model weights were trained on Minecraft textures (© Mojang/Microsoft) and community mod and resource-pack textures; they are provided for personal and community use (resource-pack creation). This project is not affiliated with Mojang.