t2i-edit-mix — geometrically repaired, bucket-aligned OmniEdit

Instruction-editing pairs derived from TIGER-Lab/OmniEdit-Filtered-1.2M, repaired so that source and target are pixel-aligned at a common resolution and shipped as webdataset shards.

This is not a mirror of OmniEdit. The upstream dataset is public and MIT; what is published here is the derivative that is expensive to recompute.

Why this exists

In upstream OmniEdit only 42.9% of pairs are pixel-aligned: src_img and edited_img disagree in size on the addition, removal and swap tasks. Any model that conditions on the source by spatial correspondence — token (i, j) of the source meaning the same place as token (i, j) of the target — needs that correspondence to hold.

The width / height columns name the canonical frame (source for addition and swap, edited for removal). The relative aspect error between the two images is small (median 1.56–2.60%, p90 ≈ 6.3%), so:

aspect error	action	share of misaligned
≤ 4%	plain resize onto the canonical frame	67–70%
≤ 12%	centre-crop to the canonical aspect, then resize	29–32%
> 12%	dropped — the crop would be inventing alignment, not correcting it	1.9%

Both images then go through the same deterministic aspect-bucket resize, so both land on one shape per pair. The invariant is verified rather than assumed: post-bucket shapes are checked per shard, and 0 mismatches occurred across 1.18M pairs.

Encoding: lossless, deliberately

Images are PNG. Source and target are encoded separately, so a lossy codec gives different values to regions that are bit-identical in the originals. Measured on natively-aligned attribute_modification pairs, JPEG q95 keeps only 54.1% of exactly-equal pixels exactly equal (q98 4:4:4 is no better at 53.5% — it is the separate-encode problem, not the quality knob). Model quality would hardly notice (48.6 dB on unchanged regions), but any exact-preservation metric would be capped at ~54% and could no longer tell a model that copies perfectly from one that does not.

Contents

task	pairs	shards	notes
swap	299,774	441	repaired
attribute_modification	250,614	357	natively aligned
addition	210,301	309	repaired
removal	210,304	309	repaired
env (environment change)	160,056	228	natively aligned, global edit
style (style transfer)	48,438	115	natively aligned, global edit
total	1,179,487	1,759	98.1% of 1,202,797 seen

Resolution: aspect buckets around 512². Shard counts are chosen so every subset has ≥ 96 shards, which a mixed webdataset loader needs (shards ≥ world_size × num_workers).

Per-shard layout

<omni_edit_id>.src.png     source image
<omni_edit_id>.tgt.png     edited image, same shape as src
<omni_edit_id>.json        {task, instruction, instructions[], o_score, sc[2], pq, src_shard, wh}

instructions is OmniEdit's full edited_prompt_list; instruction is its first entry. VIEScore columns (sc_score_1, sc_score_2, pq_score, o_score) are carried through unmodified for downstream filtering — no additional score gate was applied beyond upstream's.

Preservation is task-dependent

Fraction of pixels bit-identical between source and target, measured on these shards:

task	exact pixels	within ±2
attribute_modification	74.1%	93.1%
addition	40.6%	69.6%
removal	36.4%	68.5%
swap	32.2%	58.8%
env	0.2%	4.8%
style	0.2%	1.6%

env and style are global edits — relighting and stylisation change every pixel by design. A preservation metric averaged over all six tasks is misleading; score preservation on the local-edit tasks and instruction following everywhere.

Reproducing

omniedit/manifest.json records the resolution, per-task counts, every repair branch taken, the aspect thresholds and the image format. The generating script is handoff/editing/materialize_edit.py in the accompanying code.

License and attribution

MIT, inherited from OmniEdit. Please cite the original dataset:

@article{wei2024omniedit,
  title   = {OmniEdit: Building Image Editing Generalist Models Through Specialist Supervision},
  author  = {Wei, Cong and Xiong, Zheyang and Ren, Weiming and Du, Xinrun and Zhang, Ge and Chen, Wenhu},
  journal = {arXiv preprint arXiv:2411.07199},
  year    = {2024}
}