Inference on COCO-MIG

[Marshal874]

Dear author, thank you for your excellent work. May I ask if you could provide the inference code of the model on the COCO-MIG dataset? I first generated the reference image through FLux-Dev, but the inference result I got is quite different from the one in the paper.

[nenhang]

Thank you for your interest! The difference you're seeing is expected — the reference images used in our paper were generated with a pipeline (FLUX.1-Dev + filtering), not raw FLUX-Dev outputs. We have released the full sample pipeline and the pre-packaged reference images in bench/sample_scripts/cocomig/. You can reproduce our results directly by unpacking cocomig_reference_pack.tar.gz and running run_sample_bench_from_pack.py.
Please also note that the released open-source checkpoint is a generalized model trained across 512–1024 resolutions, so numbers will differ from the paper (which used a 512-focused checkpoint). Re-tested results with the released model are documented in bench/sample_scripts/README.md.