Bernini-Diffusers-qint8

This repository contains a qint8 quantized copy of bytedance-community/Bernini-Diffusers for Bernini full inference.

The quantized checkpoint was produced with optimum.quanto qint8 quantization. It is intended for the full Bernini pipeline only, including the MLLM semantic planner and Bernini rendering components. It does not contain Bernini-R weights and is not intended for Bernini-R inference.

Source

• Source model: bytedance-community/Bernini-Diffusers
• Quantization: optimum.quanto, qint8 precision
• Target usage: Bernini full inference
• Non-target usage: Bernini-R inference
• License: Apache License 2.0, following the source model metadata

Contents

The directory keeps the diffusers-style Bernini layout and adds the quanto quantization files for the Bernini checkpoint.

Bernini-Diffusers-qint8/
  bernini/
    model.safetensors
    quantization_map.json
    qint8_info.json
  mllm/
  scheduler/
  t5_text_encoder/
  t5_tokenizer/
  vae/
  config.json
  configuration.json
  transformer_config.json
  transformer_2_config.json

Important files:

• bernini/model.safetensors: qint8 quantized Bernini checkpoint.
• bernini/quantization_map.json: optimum.quanto quantization map required to restore qint8 weights.
• bernini/qint8_info.json: local conversion metadata.
• mllm/, t5_text_encoder/, t5_tokenizer/, vae/, scheduler/, and configuration files: supporting assets copied from the source full Bernini diffusers package.

Loading Notes

Load this directory as a local diffusers model directory in code that explicitly supports optimum.quanto qint8 restoration. The loader must read bernini/model.safetensors together with bernini/quantization_map.json before running Bernini full inference.

This package is not a replacement for Bernini-R and should not be used with Bernini-R-only inference code.