ernie-image-turbo-8bit

ernie-image-turbo-8bit

This repository contains MLX-Gen saved weights for baidu/ERNIE-Image-Turbo. The checkpoint is designed for local Apple Silicon inference with mlx-gen.

It uses the mflux/MLX saved-weight layout and MLX quantization tensors. It is not a Diffusers or Transformers from_pretrained() checkpoint.

Source Model

Original model: baidu/ERNIE-Image-Turbo.

License and Access

This quantized derivative follows the Apache 2.0 license of the source model.

Quantization

This is an MLX q8 checkpoint for ERNIE Image Turbo. MLX-Gen uses 8-bit quantization for ERNIE modules where MLX supports quantization:

• q8 for quantizable ERNIE transformer modules.
• q8 for quantizable ERNIE text-encoder modules.
• q8 for quantizable ERNIE VAE attention modules.
• BF16 for norms, convolutions, and other non-quantizable parameters.

ERNIE q4 uses a model-specific mixed q4/q8 policy because fully q4 checkpoints can drift from BF16/q8 behavior.

See the MLX-Gen quantization docs for compatibility notes and measured ERNIE q4/q8 behavior.

Prepared ERNIE folders contain the ordinary text-to-image generation stack. ERNIE Prompt Enhancer files are not bundled in this checkpoint.

Compatibility

Requires mlx-gen >= 0.18.5.

Generated with mlx-gen 0.18.5.

Use the mlxgen command and Python import path for new MLX-Gen projects.

Usage

python -m pip install -U mlx-gen

mlxgen download --model AbstractFramework/ernie-image-turbo-8bit

mlxgen generate \
  --model AbstractFramework/ernie-image-turbo-8bit \
  --prompt "Your prompt here" \
  --width 512 \
  --height 512 \
  --steps 8 \
  --guidance 1 \
  --seed 42 \
  --output image.png

Attribution

MLX-Gen is based on mflux by Filip Strand and the original mflux contributors. This model card is generated by MLX-Gen so derived checkpoints keep that attribution visible.

Quantized and contributed by @lpalbou.