PathRelax: Parallel-Path Relaxed Speculative Jacobi Decoding for Accelerating Auto-Regressive Text-to-Image Generation

PathRelax: Parallel-Path Relaxed Speculative Jacobi Decoding for Accelerating Auto-Regressive Text-to-Image Generation

This directory provides scripts for running full-dataset PathSpec evaluation with the speculative_jacobi setting and a static grouped tree (Grouped_Tree_2).

Environment

Recommended environment:

• Python 3.10
• CUDA 12.5
• PyTorch 2.5.1+cu124
• Transformers 4.47.1

Create the environment from YAML:

conda env create -f environment.yaml

Full Benchmark Evaluation Scripts

The main test entry points are the following three scripts:

• script/mscoco2017_pathspec.sh
• script/pp_pathspec.sh
• script/t2i_pathspec.sh

Before running any script, set output_path in the script to your result directory.

1) MSCOCO2017

Script: script/mscoco2017_pathspec.sh

bash script/mscoco2017_pathspec.sh

Default benchmark settings:

• prompt="MSCOCO2017Val"
• num_images=4000
• slice='1759-4000'

2) PartiPrompts

Script: script/pp_pathspec.sh

bash script/pp_pathspec.sh

Default benchmark settings:

• prompt="PartiPrompts"
• num_images=1600
• slice='1370-1600'

3) T2ICompBench

Script: script/t2i_pathspec.sh

bash script/t2i_pathspec.sh

Default benchmark settings:

• prompt="T2ICompBenchVal"
• num_images=2400
• slice='0-2400'

Notes

• mscoco2017_pathspec.sh uses ../main.py.
• pp_pathspec.sh and t2i_pathspec.sh use an absolute path to main.py. Update paths if your local directory layout is different.
• Scripts are configured with:
  • isp='random'
  • method='speculative_jacobi'
  • num_init_new_token=16
  • benchmark_way='order'
  • --static_tree --tree_choices=Grouped_Tree_2

Star History

Citation

If you use this codebase, or otherwise found our work valuable, please cite:

@misc{lei2026pathrelaxparallelpathrelaxedspeculative,
      title={PathRelax: Parallel-Path Relaxed Speculative Jacobi Decoding for Accelerating Auto-Regressive Text-to-Image Generation}, 
      author={Haodong Lei and Hongsong Wang and Bingxuan Dai and Pan Zhou},
      year={2026},
      eprint={2606.10492},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.10492}, 
}