zhenghuayu/MAPLE_40K

MAPLE_40K — 1024 × 1024

40,000 synthetic object-editing pairs derived from OBJect-3DIT / allenai/object-edit: 20,000 rotation and 20,000 translation pairs. All images are 1024 × 1024. This version replaces the previous 256 × 256 release on main.

Source and target RGB images were enhanced from native 256 × 256 using SeedVR2-7B FP16 weights, fixed seed 42, LAB color correction, and three low-frequency consistency iterations. This is model-based super-resolution. Fine details are estimates and can contain invented text or texture; these images are not native 1024-resolution renders.

This README, compressed data shards, and a repair provenance manifest are included. The release has 80 ZIP shards, 500 samples each (named data/{task}-1024-NNNNN.zip): 40 rotation shards and 40 translation shards. Download all shards for all 40,000 pairs. Each sample contains exactly five files.

Layout

Extract every ZIP into the same output directory:

rotate/
  rotate_000001/
    source.png
    rough.png
    target.png
    caption.txt
    edit_mask.png
  ...
translate/
  translate_000001/
    source.png
    rough.png
    target.png
    caption.txt
    edit_mask.png
  ...

Numbering runs from 000001 to 020000 within each task.

The rough is recomposed from source-image pixels only, using original object segmentation, fitted 2D source-to-destination geometry, and target segmentation for occlusion. Target RGB is never pasted into the rough. Premultiplied alpha is used for geometric warping.

The 256-resolution edit masks covered source support, full transformed source support, and true target support, followed by one elliptical dilation of 10-pixel radius. They are enlarged 4× with nearest-neighbor sampling, giving an approximately 40-pixel radius at 1024. Any newly changed rough support and full true target support are unioned into the enlarged mask. Masks remain exactly binary.

For pairs with sufficiently stable original backgrounds (outside-mask mean absolute RGB difference ≤ 0.5/255 and fraction of pixels with maximum channel difference >20/255 ≤ 0.001), the superresolved target is copied from the superresolved source outside the edit mask. Other targets retain the original scene changes after super-resolution. Rough pixels outside the edit mask exactly match source RGB and remain opaque.

Rotation roughs approximate 3D yaw using 2D cutouts and cannot reveal unseen object surfaces. The release retains 40,000 pairs. The 115 pairs whose edited object was invisible in both images have been replaced by unused valid official training pairs; other small, clipped, or low-motion examples remain. Captions for 39,980 examples use official asset descriptions and coarse target segmentation positions; 20 examples use visually reviewed target-scene descriptions.

Download and extract

hf download zhenghuayu/MAPLE_40K --repo-type dataset --local-dir MAPLE_40K
mkdir -p MAPLE_40K_1024
for archive in MAPLE_40K/data/*.zip; do
  unzip -q "$archive" -d MAPLE_40K_1024
done

Extract into a clean directory to avoid mixing files with a previously downloaded 256 release. When reusing a download directory, remove stale local ZIP shards from earlier releases before extracting; remote repository replacement does not delete local cached files.

from PIL import Image

rough = Image.open("MAPLE_40K_1024/translate/translate_000001/rough.png").convert("RGBA")
mask = Image.open("MAPLE_40K_1024/translate/translate_000001/edit_mask.png").convert("L")

Keep the alpha channel when loading rough images. A viewer with a black background may display transparent pixels as black; this does not change their stored alpha.

Source and attribution

Derived from OBJect-3DIT, source dataset revision 8b2be9e408c0029be14dd45867cb6ee451749c84 of allenai/object-edit. The upstream dataset lists Open Data Commons Attribution (odc-by); retain source attribution and consult the upstream project for original asset terms.

Super-resolution uses the SeedVR2-7B model with the standalone CLI implementation, vendor revision 4490bd1f482e026674543386bb2a4d176da245b9. No model weights, runtime logs, or intermediate images are distributed. The repair provenance manifest records upstream sample identifiers and file checksums.

Invalid-case repair (2026-10-08)

115 unusable pairs (107 rotation, 8 translation) were replaced under the same sample IDs with unused, numerically screened official training pairs. Their original edited-object segmentations were empty in both source and target; no valid visible edit could be reconstructed. Source, target, rough, binary edit mask, and target-scene caption are updated together. No masks were fabricated to make the validator pass.

Replacements have source/target object support of at least 512 native pixels, visible planar destination IoU at least 0.85, and stable backgrounds. Their 1024 rough images use nearest-neighbor source-object support and premultiplied affine pasting so vacated regions remain transparent. Target RGB is never pasted into rough. Template captions and the 2D approximation of 3D yaw still require scientific quality review; these checks do not establish benchmark performance.

Per-case provenance and file checksums record the previous and replacement upstream IDs. Exactly 40,000 unique sample folders and the existing shard layout are preserved. Unaffected ZIP members were verified by SHA-256. Rebuild training indices after downloading this update, and extract into a clean directory.

Resolution status: All 40,000 published pairs are 1024 x 1024. The 115 repaired samples and their provenance are preserved. The complete server-side dataset passes the strict preparation script with 40,000 accepted and zero rejected cases.