Track4D-2026

Public checkpoints and exact evaluation assets for Track4D, a conditional-video approach to dense 3D tracking.

Files

• ncmv/wan21_t2v_14b_512x512x49_lora_r64_step4000.safetensors: NCMV-only Wan2.1-T2V-14B adapter, native 512 x 512 x 49.
• rtv/joint_rtv_ncmv_wan21_i2v_14b_lora_r64_step5000.safetensors: joint RTV+NCMV Wan2.1-I2V-14B adapter, trained with 49-frame diverse-resolution data.
• rtv/rtv_only_noqi_wan21_i2v_14b_lora_r64_step9000.safetensors: RTV-only/noQI Wan2.1-I2V-14B adapter, trained with 49-frame diverse-resolution data.
• experimental/da3_vdm_vitdpt_step016000/: retained Wan LoRA, ViT/DPT probe, and DA3 LoRA for the active live-DA3 3D tracking pipeline.
• experimental/dense2d_step013500/: retained Wan LoRA and ViT/DPT probe for dense UV tracking plus visibility.

These are adapters, not standalone Wan checkpoints. Download the matching Wan2.1 base model separately. Exact SHA256 hashes, inference configs, training configs, and WorldTrack evaluation commands are in the code repository.

For the released WorldTrack scorecard, short256 is both the aligned membership and the short-side-256 Stream3R scaffold contract. ADT uses 256 x 256; PO, PStudio, and DS/DR preserve landscape aspect and use the 448 x 256 model bucket. Joint RTV+NCMV uses one denoising step; RTV-only uses 50. Exact replay uses base seed 42 plus the global video index over the fixed ADT, PO, PStudio, DS/DR order.

The joint RTV+NCMV model is a continuation from RTV-only. Its historical score improvement describes the complete two-stage recipe and is not an isolated causal estimate of the NCMV loss.

Evaluation assets

The code repository downloads each group selectively with scripts/download_eval_assets.py:

• evaluation/worldtrack_release/: 208 benchmark input/GT NPZ files.
• evaluation/stream3r_short256_49f/: 200 aligned RGB/geometry NPZ files for the stable RTV replay.
• evaluation/da3base_step012500_predictions/: exact historical 3D predictions plus evaluator JSON/TXT.
• evaluation/dense2d_step011500/: five-source metric JSONs and qualitative NPZ/MP4 examples.

The old experimental 3D step-12.5K and dense-2D step-11.5K model weights were rotated during continued training. Their saved outputs reproduce the published numbers exactly; the retained 16K/13.5K weights are runnable but are not misidentified as the historical metric checkpoints.

Dataset-specific redistribution terms remain separate from the Track4D code license. A few provenance-only string fields in historical NPZ files record the original cache path; evaluation uses only the embedded arrays.