3DEdit-1M

A large-scale 3D edit-pair dataset of the paper Omni123: Exploring 3D Native Foundation Models with Limited 3D Data by Unifying Text to 2D and 3D Generation.

Forward edit pairs	1,476,827 (487,928 distinct source objects)
Reverse edit instructions	1,356,826 (91.9% of the pairs)
Shards	211 WebDataset tars, data/shard-{000136..000346}.tar
Size	309.7 GB

Each sample is a paired (source, target) 3D edit. The pipeline behind every sample:

• Edit instruction synthesized by Qwen3.5-35B-A3B.
• Target image (post-edit RGB render) generated by FLUX.2-klein-9b-kv.
• Source and target 3D meshes generated by Hunyuan3D-2-mini from the source and target images.
• Shape tokenization: each mesh is tokenized with the cube3d v0.5 shape encoder (OneDAutoEncoder); the released .npy files store those discrete encoder indices.

Dataset structure

data/
├── shard-000136.tar
├── shard-000137.tar
├── ...
└── shard-000346.tar          # 211 shards, 7,000 samples each (last one 6,827)
reverse_instructions.jsonl    # reverse edit instructions, keyed by sample id

Sample keys are <uuid>_<k>: a single source object (uuid) yields up to 4 different edits, indexed by k.

File	Description
<id>/<id>.source.npy	Source shape token indices from the cube3d v0.5 encoder — shape (1, 1024) int64, codebook size 16,384 (value range [0, 16383])
<id>/<id>.target.npy	Target shape token indices — same encoder, same shape/dtype
<id>/<id>.source.rgba.webp	Source RGBA render, 1024×1024, lossless WEBP
<id>/<id>.target.rgba.webp	Target RGBA render, 512×512, lossless WEBP
<id>/<id>.meta.json	Edit instruction and metadata (see below)

meta.json fields

{
  "id": "b734f7f2-62f3-4397-882c-98f9387f66bc_1",
  "uuid": "b734f7f2-62f3-4397-882c-98f9387f66bc",
  "captions": ["long caption ...", "medium caption ...", "short caption"],
  "instruction": "Replace the current side mirror with a smaller, more angular, and aerodynamically shaped mirror."
}

captions describe the source (pre-edit) object at three levels of detail. instruction is the edit that was actually applied to produce this sample's target.

Loading with WebDataset

import webdataset as wds

url = "https://huggingface.co/datasets/meshy-ai-team/3DEdit-1M/resolve/main/data/shard-{000136..000346}.tar"
ds = (
    wds.WebDataset(url, shardshuffle=True)
    .decode("rgb")
    .to_tuple("source.npy", "target.npy",
              "source.rgba.webp",
              "target.rgba.webp",
              "meta.json")
)

Reverse edit instructions

The shards contain forward edits only (source → target). reverse_instructions.jsonl supplies the inverse instruction for each sample, so a reverse pair can be built by swapping source and target and using the reverse instruction in place of meta.json["instruction"]. Together the two directions give ~2.83M edit pairs.

The reverse instructions were generated by Qwen3.6-27B (thinking mode, dual-image): the model is shown the edited object and the original object together with the forward instruction, proposes candidate inverse instructions, filters out the ones that are colour-only, scale-only or vague, and selects the best. Grounding the model in both images lets the inverse name concrete properties of the original part that the forward instruction never mentions:

forward : Replace the tail rotor assembly with a larger, more angular multi-blade tail rotor.
reverse : Replace the large multi-blade tail rotor with a smaller two-bladed tail rotor assembly.

One JSON object per line:

{"id": "0000661e-710b-4e07-9ec4-571531a94804_0",
 "uuid": "0000661e-710b-4e07-9ec4-571531a94804",
 "instruction": "Replace the large multi-blade tail rotor with a smaller two-bladed tail rotor assembly.",
 "fallback": false}

Field	Description
id	Matches the sample key inside the shards — join on this
uuid	Source object id (the id without the _<k> suffix)
instruction	The reverse edit instruction
fallback	true when no candidate survived filtering and a best-effort pick was used (117,503 rows, 8.66%). Lower quality — filter these out if you need a clean subset.

Notes on coverage:

• 1,356,826 of the 1,476,827 samples have a reverse instruction (91.9%). The remaining 120,001 were never generated or were dropped as degenerate outputs; join on id and skip misses rather than assuming full coverage.
• Instruction length ranges from 20 to 497 characters (median 69).

License

Released under the Apache License 2.0.