Rethink Benchmark

Unified 13-route benchmark. Every route is one underscore-normalized named split in the default config, backed by one root-level Parquet file with 100 records.

Shared schema

• idx: numeric row index from 1 to 100.
• mode: route-specific benchmark mode/category.
• prompt: generation or editing instruction.
• target: target/reference output image when the benchmark provides one; otherwise null.
• input_1 ... input_10: ordered conditioning images; unavailable positions are null.

Ten input slots are required by the largest public MI2I benchmark. I2I_HELDOUT uses input_1 for its source image. MI2I_HELDOUT preserves reference order in input_1 ... input_4 and leaves unused positions null. Images are embedded directly as Hugging Face Image features.

All five T2I splits are prompt-only evaluations, so target and every input_* column are null. Qwen-Image-Bench includes candidate generations from multiple evaluated models, but it does not designate any of them as ground truth; those model outputs are therefore not projected as target.

Hugging Face split identifiers cannot contain dashes, so split names use underscores while the root-level Parquet filenames retain the project route names.