Inference Time Scaling for Flux T2I

This repository implements the ["Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps"] (https://arxiv.org/html/2501.09732v1). We only focus on the third search algorithm, "search over paths" in this repository.

Run the code

Set the environment variable OPENAI_API_KEY if you need to use OpenAI verifier.

Install the deps in requirements.txt and then execute:

python run_search_over_path.py

Notes

1. The verifier(scorer)'s implementation is following the pattern from https://github.com/sayakpaul/tt-scale-flux. We use OpenAI endpoint. You may use other verifiers such as HPSv2 etc.
2. We customized flux pipeline and scheduler to allow both step_back operation and linear-extrapolation denoising to x0 from a high noise level. We need this estimated x0 for verifier to make precise decisions.