---
configs:
- config_name: default
  data_files:
  - split: test
    path: data_bench_v1/test-*
  - split: real_photo
    path: real_photo/0000.parquet
---

# Material Replacement Bench v1.0

Frozen evaluation split for the interior-design material-replacement
benchmark: 1421 task specs (mask-aligned, watermark-free), of which
1093 are `gt_clean` (the stored `output` edits only the mask and
matches the reference; the rest keep `output` as a non-exemplar reference
solution only).

Task: given `input` + `mask` + `reference` (or the mask-free `instruction`),
replace the masked object's material and change nothing else. Evaluation is
reference-free — outputs are scored on material correctness, outside-mask
preservation, geometry preservation, and lighting harmony, never on pixel
similarity to `output`.

Columns: `instruction` (rewritten, image-faithful), `material_category`,
`surface_type`, `mask_target_seen`, `mask_coverage`, `coverage_bucket`,
`gt_clean`, plus the four images.

Provenance: fully synthetic quadruples generated with OpenAI image models
via codex; curated from `Elfsong/vinhome_material_replacement` (algorithmic
QC + VLM semantic QC + instruction relabeling; see `curation/` there).
Frozen 2026-07-20. Research use.