pipeline_tag: text-to-image library_name: diffusers license: apache-2.0 tags:

• pixel-art
• game-development
• aseprite
• pico-8
• sprites
• cli
• mlx
• windows
• cuda
• apple-silicon
• local

Bilta Gilata1

Turn a sentence into palette-locked pixel art—locally on macOS or Windows.

Bilta Gilata1 is a free local pixel-art CLI for Apple Silicon and NVIDIA-powered Windows PCs. It has 65 Aseprite palettes, reference-image guidance, deterministic seeds, and exact 16–256 px output. No paid API or uploaded prompts.

Why use it?

• Actually local: MLX runs on Apple Silicon; PyTorch/CUDA runs on Windows.
• Exact palettes: output is quantized to one of 65 Aseprite palettes.
• PICO-8 ready: includes the official 16-color palette option.
• Reference images: guide composition and character appearance from a local file.
• Reproducible: reuse --seed to regenerate a composition.
• Game-sized: export at 16, 32, 48, 64, 128, or 256 pixels.

Install on Apple Silicon

git clone https://github.com/Abd196-bit/bilta-gilata1.git
cd bilta-gilata1
chmod +x install.sh
./install.sh

The installer downloads the model into Gilata1's local cache, so generation works offline afterward. The model is not placed inside the Python wheel because the Windows checkpoint alone is approximately 24 GB.

Install on Windows 10/11

Requirements: Python 3.10+, a current NVIDIA driver, an NVIDIA GPU, roughly 24 GB RAM, and about 25 GB free disk space for the first model download. More GPU VRAM is faster; CPU offload is enabled for smaller cards.

Open PowerShell in the cloned folder:

git clone https://github.com/Abd196-bit/bilta-gilata1.git
cd bilta-gilata1
Set-ExecutionPolicy -Scope Process Bypass
.\install-windows.ps1
.\gilata1.bat "a tiny knight" --palette pico-8 --size 64 --output knight.png

If the installer reports that CUDA is unavailable, use the selector at pytorch.org/get-started/locally to install the CUDA build matching your NVIDIA driver.

Generate

.venv/bin/bilta-gilata1 "a tiny moon wizard" --palette pico-8 --size 128 --output wizard.png

Reference-guided generation:

.venv/bin/bilta-gilata1 "a forest adventurer" \
  --reference portrait.jpg --palette db16 --size 64 --output adventurer.png

Browse palettes:

.venv/bin/bilta-gilata1 --list-palettes

Python library

The CLI and Python API use the same generator:

from bilta_gilata1 import generate

result = generate(
    "a tiny moon wizard",
    output="wizard.png",
    palette="pico-8",
    size=64,
    seed=1234,
)
print(result.output, result.seed, result.backend)

Reuse project defaults and the loaded CUDA pipeline across images:

from bilta_gilata1 import Gilata1, list_palettes, palette_colors

artist = Gilata1(palette="db16", size=64)
artist.generate("a walking knight", output="knight.png")
artist.generate("a moon wizard", output="wizard.png", reference="sketch.png")

print(list_palettes())
print(palette_colors("pico-8"))

generate() returns a GenerationResult containing output, master, seed, backend, palette, and size. Pass on_progress=print to receive progress messages in an application.

Generate a complete batch with one call. On CUDA, the model is loaded once and reused for every prompt:

from bilta_gilata1 import generate_bulk

batch = generate_bulk(
    [
        "a walking knight",
        "a moon wizard",
        "a cyberpunk fox merchant",
    ],
    output_dir="game-assets",
    palette="pico-8",
    size=64,
    seed=500,
    continue_on_error=True,
    on_progress=print,
)

print(batch.outputs)
print(batch.failures)

The output names are deterministic, such as 001-a-walking-knight.png. With seed=500, the batch uses seeds 500, 501, and 502.

Pre-download the model for a fully offline machine:

from bilta_gilata1 import download_model

download_model(on_progress=print)

Bulk generation is also available from the CLI. Put one prompt on each line:

bilta-gilata1 --bulk prompts.txt --output-dir game-assets --palette pico-8 --size 64 --seed 500
bilta-gilata1 --download-model

Requirements

• Apple Silicon Mac, or Windows 10/11 with an NVIDIA CUDA GPU
• Python 3.10+
• About 5 GB free on Apple Silicon or 25 GB on Windows for the first model download
• Internet required only for the initial download

Gallery

Sprite	Animation study	Palette output

License and responsible distribution

Bilta Gilata1's original source code is licensed under Apache-2.0. This repository distributes the runtime and palette post-processor, not third-party model weights. Models, adapters, and palette files remain subject to their own terms; see THIRD_PARTY_NOTICES.md.