Elfsong/material_replacement_submissions

Real-photo C+ supplement

gpt-image-2-Cplus, split real_photo, contains the 100 first-returned GPT-Image-2 outputs for the fixed real_photo tasks. C+ supplies the text instruction, original scene, binary mask, and material reference. Prompts and input order match the user-confirmed Nano-Banana configuration. Native output sizes are preserved.

Source tasks: Elfsong/material_replacement_bench, revision 01f5bec3f1161df278ce4e4a03f61174dd54c7fe. Join by sample_uid. Generation protocol, exact prompts, hashes and source-review observations are stored under gpt-image-2-Cplus/provenance/real_photo/. The split name is inherited; this run does not independently certify all inputs as natural photographs.

Open-model real_photo C+ supplement

qwen-image-edit-2509-Cplus, omnigen2-Cplus, and bagel-Cplus each contain 100 outputs for the same fixed real_photo tasks. All use the original scene, mask, material reference, and the same full text instruction (C+). Official model revisions, inference settings, exact prompts, hashes, and first-output audits are included per configuration. No candidate selection or background post-compositing is applied.

The five-model paired evaluation is in real_photo-Cplus-evaluation/summary.json. It evaluates the three open models plus the existing GPT-Image-2 and Nano-Banana outputs using the same pinned CUDA metric models. The primary Gate retains the paper's fixed calibration; the separately labeled sensitivity result retains the previously established real_photo anchors. Previous CPU reports remain available at their original paths.

These are C+ results, not the paper's A/B/C tracks. The original paper's exact open-model checkpoint hashes were unavailable for verification. Qwen and BAGEL output canvases explicitly follow the original scene; upstream multi-image inference otherwise inherits the last reference image's canvas. One Qwen interface-validation pilot used the incorrect reference canvas and was excluded before the uniform final protocol; this is documented in its preflight note. Source provenance observations remain documented; no tasks were selectively removed.

Rebuttal P0 supplement

Eight configurations ending in -rebuttal-P0 contain 1,200 new outputs. Qwen-Image-Edit-2509 and OmniGen2 each have 100 synthetic tasks under B (scene, mask, text), C (scene, mask, reference; neutral instruction), and C+ (all conditions), plus 100 real-photo C tasks. ZeST and MaterialFusion each have 100 synthetic and 100 real-photo C tasks. Ten independent interface-pilot tasks per specialist are excluded from the formal tables.

The balanced synthetic sample is a diagnostic subset, not a population-proportional estimate. The same pinned checkpoints, one fixed seed, and first valid output are used throughout each method. Specialist square-canvas processing and resizing back to scene dimensions are disclosed in the per-run protocols.

All formal P0 generations use B200/CUDA 12.8. The 200 completed H100 Qwen outputs are archived and excluded; Qwen B/C/C+ are regenerated together to keep their runtime fixed. All 1,900 scores, including the existing five-model C+ images and Identity/Paste anchors, are recomputed on B200 with the same pinned metric models and fixed calibration. Six independent hardware-preflight outputs per editor are excluded. The previously completed real-photo C+ generations of Qwen and OmniGen2 used H100, so a comparison against their new real-photo C generations also changes generation hardware/runtime. The historical H100 evaluation remains available separately. MaterialFusion clones fixed prompt embeddings outside inference mode to permit gradient checkpointing; the compatibility wrapper is included, and the initial zero-output failed interface attempt is documented in the implementation notes.

Summary and paired confidence intervals keep C and C+ separate, retain the fixed Gate calibration, and add Identity/Paste anchors to the five-model real-photo C+ evaluation. Intervals are descriptive and unadjusted for multiple comparisons. Source and mask audit distinguishes automatic checks from pending source verification and human mask review; source-dataset metadata alone does not establish photographic provenance.

Real-photo Track A supplement

Five configurations ending in -A-real-photo contain 500 outputs on the same 100 tasks. Track A supplies only the original scene and exact task instruction; the generator receives neither a mask nor a material swatch. The three open models use the pinned H100 environments from the completed real-photo C+ runs. Nano-Banana uses the Gemini REST model ID gemini-3.1-flash-image-preview, with the returned model version recorded per sample. GPT-Image-2 uses the Codex built-in image generator, with no exposed snapshot, seed, or quality controls, matching the backend of the earlier real-photo C+ run. Seven independent interface-check outputs (six open-model and one Nano) are excluded. The existing gpt-image-2-A/test dataset has 1,421 task IDs with zero overlap with these 100 real-photo tasks; its outputs and the existing real-photo C+ outputs are preserved.

All 500 outputs are scored with the same B200 metric environment, metric checkpoints, code, and fixed Gate calibration as the P0 supplement. The summary includes task-paired bootstrap intervals, the existing B200-scored Identity/Paste anchors, and within-model A versus C+ differences. Masks and swatches are used by the evaluator only. Historical API snapshots/default generation settings and exact original-paper checkpoint matches are not independently verified. P0 real C generation uses B200, which also differs from the H100 generation environment of these open-model A outputs. Source verification and human mask review retain the pending status reported in the P0 source audit.

GPT-Image-2 real-photo B/C supplement

Two configurations, gpt-image-2-B-real-photo and gpt-image-2-C-real-photo, add 100 outputs each on the same 100 real-photo tasks. B supplies the original scene, binary mask, and exact target-material text; C supplies the original scene, mask, and reference swatch with one neutral instruction and no target-material text. Existing A and C+ outputs are reused. All four conditions use the Codex built-in GPT-Image-2 backend, whose exact model snapshot, seed, and quality controls are not exposed. Each task retains its first successful native output without aesthetic retries or postprocessing.

The four-condition report contains 400 scored rows, including the 200 new B/C images, and paired-task bootstrap intervals. Scoring uses the same pinned B200 evaluator and fixed Gate calibration as the A/P0 reports. A has a shorter prompt than B/C/C+, so A-versus-B measures a protocol change; provider snapshot consistency across run times is unverified. Source and human mask review remain pending, and observed text-mask discrepancies are retained and documented without selective exclusions.

Seed and input-format robustness supplement

800 new outputs from Qwen-Image-Edit-2509 and OmniGen2 cover a fixed, score-blind 50-task synthetic subset. A/B use three seeds; reference conditioning additionally tests explicit image-role wording and reordered images with scene canvas preserved. Qwen reuses 100 original B0/C0 outputs after exact pilot reproduction. OmniGen2 regenerates these baselines in the same process as all variants because separate loading processes did not reproduce exactly. No aesthetic selection or post-compositing is used.

Paired results and limitations distinguish task uncertainty from the limited three-seed diagnostic. Metric validation does not establish that deleting every dimension hurts human alignment. New two-rater overlap, per-dimension judgments and human source/mask audits remain pending.