Unlike pixel-art style LoRAs, the output is not "an illustration that looks like pixel art". It is an actual 16×16 RGBA sprite you can drop into a resource pack.

The one thing you must know

Generate at 256×256, then fold each 16×16 block into one pixel.

The model was trained exclusively on 16×16 textures upscaled ×16 with nearest-neighbour. Every training image therefore consists of 16×16 flat blocks of 16×16 pixels each, so the model only ever draws on that grid. Folding the 256px output by block-averaging recovers the original 16×16 sprite with no loss.

Transparency is encoded as magenta (255, 0, 255) and keyed out afterwards. That colour was chosen because only 0.057% of opaque pixels in the training data come near it (black would collide with outlines in 23% of images, white in 25%).

Use an empty negative prompt. Training used caption dropout to an empty string, so the empty string is the model's unconditional path. Non-empty negatives contain words that are almost absent from the 11,613-word training vocabulary, and pushing away from them injects noise. Measured: with the usual "blurry, photo, watermark…" negative, usb drive and cat never resolve at any guidance scale; with an empty negative both work.

Usage

python

import numpy as np, torch
from PIL import Image
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler

BG = np.array([255, 0, 255], dtype=np.int16)   # chroma key colour

pipe = StableDiffusionPipeline.from_pretrained(
    "yuuki14202028/minecraft-item-16px", torch_dtype=torch.float16,
    safety_checker=None, requires_safety_checker=False).to("cuda")   # or "mps"
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

image = pipe("copper sword", negative_prompt="", num_inference_steps=30,
             guidance_scale=9.0, height=256, width=256).images[0]

def to_tile(image, size=16, tol=100):
    a = np.asarray(image.convert("RGB"), dtype=np.float32)
    c = a.shape[0] // size
    rgb = a.reshape(size, c, size, c, 3).mean(axis=(1, 3)).round().clip(0, 255).astype(np.int16)
    alpha = np.where(np.abs(rgb - BG).sum(axis=2) <= tol, 0, 255).astype(np.uint8)
    return Image.fromarray(np.dstack([rgb.astype(np.uint8), alpha]))

to_tile(image).save("copper_sword.png")        # 16x16 RGBA

Recommended settings: 256×256, guidance 9.0, 30 steps, empty negative prompt, DPM-Solver++. Guidance 7–12 all work; 3–5 is visibly mushy.

What it can and cannot do

Measured on 50 prompts × 2 seeds and a 20-prompt × 3-seed diagnostic set.

Strong. Material × form composition is essentially perfect, including pairs that never co-occur in the training data: glass hammer (glass and hammer appear separately, never together) and copper crown both hit 3/3. diamond crown, emerald key, gold skull, glass sword — 16/16.

Good. Objects absent from Minecraft: coffee cup, donut, cheese, syringe, gear, umbrella, sushi, penguin, jellyfish, torii gate, katana. Abstract prompts like black hole and rainbow also produce sensible sprites. Multi-adjective prompts (glowing molten iron sword, ancient mossy stone tablet) mostly land.

Weak. Concepts with zero training-vocabulary support rely entirely on the SD1.5 prior and hit roughly 2 in 3 (guitar, microscope). Fine-grained modifiers such as "tiny" and "rusty" are often ignored. Around 10–15% of generations fill the whole frame instead of leaving a background, so the chroma key finds nothing — regenerate with another seed.

Judge by hit rate, not by one image. Seed variance is large. Single samples repeatedly gave the wrong impression during development: steel ingot looked broken until 4 seeds showed 3/4 correct, and banana looked impossible until 6 seeds showed 5/6 correct.

Training


Base	`stable-diffusion-v1-5/stable-diffusion-v1-5`, UNet fully unfrozen (859.5M)
Frozen	VAE (83.7M), CLIP text encoder (123.1M)
Data	369,342 item textures, 16×16 RGBA, upscaled ×16 to 256px
Captions	derived from file names (`copper_sword.png` → "copper sword"), 11,613-word vocabulary
Steps	92,335 (4 epochs, batch 16), lr 1e-5 cosine, 500 warmup
Objective	epsilon prediction with min-SNR weighting (γ=5), 10% caption dropout
Hardware	1× RTX 4090, 6.2 hours, 66.8 img/s

Loss over the run: 0.01192 → 0.00858, flattening in the final quarter as the cosine schedule decayed, which suggests 4 epochs was about the right length for this configuration.

Training data provenance and licensing

Trained on OVAWARE/16xModdedMinecraft, a gated dataset of 1,034,057 textures scraped from Minecraft mods on Modrinth. Only the 552,357 entries labelled item were used, reduced to 369,342 after removing exact duplicates, near-transparent and single-colour images, UI elements, and entries whose file names yielded no usable caption.

The dataset carries 235 distinct license values, including LicenseRef-All-Rights-Reserved. The weights are not a copy of any texture, but generated outputs may resemble existing mod assets. If you redistribute generated textures — in a resource pack, a mod, or a commercial product — verify their originality yourself. This model is released for research and personal use; the authors of the original textures did not consent to this specific use.

Base model weights are CreativeML OpenRAIL-M and that license carries over here.

Limitations

Items only. Full-bleed block textures were deliberately excluded from training.
16×16 only. There is no meaningful detail to extract at higher folds.
English prompts only, and the caption vocabulary is skewed toward modded Minecraft terms.
No EMA was used during training; adding it would likely improve sample quality.
The text encoder was left frozen. Unfreezing it at a low learning rate is the most obvious next experiment for prompt adherence.

Code

Training, sampling, evaluation and img2img tooling: github.com/yuuki14202028/pixelgen

日本語

Stable Diffusion 1.5 のUNetを、Minecraft MODのアイテムテクスチャ369,342枚でフルファインチューンしたモデルです。 出力は「ピクセルアート風の絵」ではなく、リソースパックにそのまま入る本物の16×16 RGBAテクスチャです。

使い方の要点は3つ。256×256で生成し、16×16画素ずつ平均して畳むこと。**透過はマゼンタ(255,0,255)**で表現されているのでキーイングして戻すこと。そして negative promptは空にすること（学習時の無条件経路が空文字列なので、非空のnegativeはノイズになります）。

推奨設定は 256×256 / guidance 9.0 / 30ステップ / negative空 / DPM-Solver++ です。