Reproduce “1-Bit Quantization of FLUX.2 Klein 4B Disproportionately Hurts Compositional Benchmark Scores” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Avaya Aggarwal
The authors audit whether one-bit quantization preserves prompt following in FLUX.2 Klein 4B, comparing it with a full-precision reference under shared generation and scoring conditions. On 553 GenEval prompts at 512 pixels, the one-bit model scores 0.645 versus 0.824, with the largest losses for counting and position. This benchmark result does not establish general image quality or faster generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.