Causal-Adapter

Causal-Adapter Pretrained Weights

Model Overview

This repository provides pretrained Causal-Adapter weights across four benchmark settings. The released checkpoints include Causal-Adapter models built on both SD1.5-style and SD3-style diffusion structures.

Causal-Adapter is designed to inject structured causal semantics into pretrained text-to-image diffusion models for controllable and causally consistent counterfactual image generation.

Detailed usage examples are available in our notebook benchmarks:

Notebook Benchmarks

An example configuration can be found in:

notebook_benchmarks/counterfactuals_celeba.ipynb

Base Models

The released checkpoints are based on the following pretrained diffusion backbones:

• SD1.5-style structure: lambda/miniSD-diffusers
• SD3-style structure: stabilityai/stable-diffusion-3-medium-diffusers

Benchmark Resources

The released weights are evaluated on benchmark settings built from the following resources:

• Pendulum dataset generation:
  CausalVAE Pendulum

• CelebA and ADNI benchmark configuration:
  counterfactual-benchmark

• CelebA-HQ dataset:
  CelebAMask-HQ

Example Configuration

The following example shows the main paths required for running the CelebA counterfactual generation notebook.

import os

# Shared roots

# 1) Frozen SD1.5 backbone.
# For example: "lambda/miniSD-diffusers"
BASE_MODEL_PATH = ""

# 2) Causal-Adapter ControlNet checkpoint and the matching MCPL learned pseudo-tokens.
# Example ControlNet checkpoint:
# https://huggingface.co/LeiTong/Causal-Adapter/tree/main/celeba/controlnet/controlnet-steps-200000.safetensors
CONTROLNET_PATH = ""

# Example learned text embeddings:
# https://huggingface.co/LeiTong/Causal-Adapter/tree/main/celeba/controlnet/learned_embeds-steps-200000.safetensors
TEXT_EMBEDDING_PATH = ""

# 3) Optional pretrained SCM head from SCM_modeling/.
# Example SCM checkpoint:
# https://huggingface.co/LeiTong/Causal-Adapter/tree/main/celeba/scm/best_model.pt
SCM_PATH = ""

# 4) CelebA root expected by torchvision.datasets.CelebA(root=...).
DATA_ROOT = os.environ.get("DATA_ROOT", "")

DATASET = "celeA_complex"
SIZE = 256

Repository Structure

The checkpoint files are organized by benchmark and model component. A typical setting may include:

• Causal-Adapter / ControlNet weights
• Learned pseudo-token embeddings
• Optional pretrained SCM head
• Example notebooks for counterfactual image generation

Usage

Please refer to the notebook examples for loading the pretrained weights and running counterfactual generation:

Notebook Benchmarks

License

This repository is released under the Apache-2.0 license.