Practical overview
When it fits
Use MACE when you want to investigate or implement multi-concept erasure in a text-to-image diffusion model, rather than use a hosted generation app.
Who it is for
- Researchers studying concept erasure and unlearning in diffusion models
- Practitioners preparing and fine-tuning text-to-image models to suppress selected concepts
Documented uses
- Prepare data and masks for concept-erasure fine-tuning
- Train a diffusion model to suppress selected concepts in generated images
- Sample from a modified model and evaluate erasure metrics
Requirements & limits
- The README lists a Python 3.10 environment with PyTorch 2.0.1, torchvision 0.15.2 and CUDA 11.7.
- Data preparation uses Grounded-SAM; the README says its official version can run on a 24GB GPU, while the Hugging Face version likely requires above 28GB.
- Text augmentation uses an LLM. OpenAI with gpt-3.5-turbo is the default; MiniMax can be selected through environment variables.
- The documented workflow includes preparing data, training, sampling from the modified model and evaluating metrics.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up the core MACE environment
Create the Conda environment and install the README’s listed core model dependencies.
Before you start
- Conda
- Python 3.10
- PyTorch 2.0.1, torchvision 0.15.2 and pytorch-cuda 11.7, installed from the PyTorch and NVIDIA Conda channels
Steps
Clone the MACE repository.
Create and activate a Conda environment named mace with Python 3.10.
Install the listed PyTorch, torchvision and CUDA packages.
Install Diffusers 0.22.0, Transformers 4.46.2, Hugging Face Hub 0.25.2, Accelerate, OpenAI, OmegaConf and OpenCV-Python.
Expected outcome
The repository’s documented core environment and Python dependencies are set up; the README describes data preparation and training as later stages.
Guide limitations
- This setup recipe does not cover Grounded-SAM installation, data preparation, training or sampling.
- The README presents separate Grounded-SAM preparation options and gives different GPU-memory guidance for them.
- Text augmentation requires configuring an LLM provider and its API key.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- Shilin-LU/MACE · Official homepage title · retained in its source language