BiGain: Unified Token Compression for Joint Generation and Classification
March 19, 2026 ยท View on GitHub
Official implementation of the CVPR 2026 paper BiGain: Unified Token Compression for Joint Generation and Classification.
Framework of our BiGainTM method. A Laplacian filter is applied to hidden-state tokens to compute local frequency scores. In each spatial stride, the lowest-scoring token is selected as a destination token, while the others form the source set. Destination and source tokens are gathered globally, and a bipartite matching selects top source-destination pairs.
Environment Setup
- Create conda environment:
conda create -n diffusion-exp python=3.9
conda activate diffusion-exp
- Install dependencies:
pip install -r requirements.txt
- Install ToMe package:
cd tomesd && python setup.py build develop && cd ..
- Set dataset path:
export DATASET_ROOT=/path/to/your/datasets
Running Experiments
All experiments are configured through shell scripts in the scripts/ directory.
Classification Experiments
# Stable Diffusion zero-shot classification
bash scripts/sd_classification.sh
# DiT (Diffusion Transformer) classification
bash scripts/dit.sh
Generation Experiments
# Stable Diffusion image generation
bash scripts/sd_generation.sh
# DiT image generation
bash scripts/dit_generation.sh
Configuration
To run an experiment:
- Open the desired script
- Uncomment one of the example configurations
- Adjust parameters if needed
- Run the script
Available Methods
- Baseline: No acceleration
- ToMe: Original Token Merging
- BiGain_TM (LGTM): Our scoring-based token merging method
- ToDo: Token downsampling
- BiGain_TD (IEKVD): Our linear blend token downsampling method
- SiTo: Similarity-based token pruning
Code Attribution
This implementation builds upon code from:
- Zero-shot classification framework: Li et al., "Your Diffusion Model is Secretly a Zero-Shot Classifier", ICCV 2023 [Paper] [Code]
- ToMe: Bolya & Hoffman, "Token Merging for Fast Stable Diffusion", CVPR Workshops 2023 (MIT License) [Paper] [Code]
- ToDo: Smith et al., "ToDo: Token Downsampling for Efficient Generation of High-Resolution Images", IJCAI 2024 [Paper] [Code]
- SiTo: Zhang et al., "Training-Free and Hardware-Friendly Acceleration for Diffusion Models via Similarity-based Token Pruning", AAAI 2025 [Paper] [Code]