va1bhavagrawa1/data_mix_v4_3d_wireframes

Data Mix v4

This is a 60,000-record paired image-editing dataset packaged as 1,200 tar shards (50 records per shard). manifest.jsonl indexes each record and its six members: reference RGB, target RGB, reference spatial map, edited spatial map, edited depth map, and depth metadata.

Source	Records
Data Mix v2 ST3D four-object movement	10,000
Data Mix v2 PaintBench four primitives	10,000
Data Mix v2 PartNet Mobility articulation	10,000
Corrected Data Mix v3 textured four-object scenes	10,000
Additional copies sampled from that v3 four-object category	5,000
Corrected Data Mix v3 single-object rotation	10,000
Additional copies sampled from that v3 single-object category	5,000

The 10,000 additional v3 records intentionally repeat existing v3 examples. Each has a unique v4 ID and a source_sample_id identifying its original. Sampling is deterministic with seed 20261024.

ST3D object movement and PartNet articulation use projected 3D box or primitive wireframes on gray RGB backgrounds. Their map colors track reference image position and camera depth. PaintBench uses 2D primitive outlines on the same gray background; its blue boundary channel is fixed and its edited depth is ordinal layer depth. The two v3 categories come from the corrected RGB-box dataset, with true edited depth maps.

The source code for packaging is diffusion_training/build_data_mix_v4.py in the spatiallang workspace. See config.json for the exact counts and source directories used to build this release.