We study how to repurpose video diffusion models (VDMs), pretrained on millions of videos, for dense 3D point tracking, where training data is scarce. Doing so is non-trivial, as it requires representing 3D tracks as video-like signals that may lie outside the natural-video distribution on which the VDM was trained. We thus systematically assess the effectiveness of two such video representations. The first, Residual 3D Tracking Video (RTV), encodes per-pixel 3D offsets relative to a query frame and allows recovery of 3D trajectories directly from the generated video. The second, Normalized Coordinate Map Video (NCMV), labels each pixel with a canonical coordinate invariant through time. This is closer to a natural video and easier for the VDM to learn, but requires additional post-processing with an off-the-shelf 2D tracker to recover 3D tracks. RTV tracks more accurately, while NCMV shows that a VDM can generate dense correspondence as a video. We instantiate both representations with LoRA adaptation of a frozen Wan backbone, conditioned on point maps from an off-the-shelf 3D reconstructor as a geometric scaffold. Trained with a single dynamic 3D tracking dataset, Track4D achieves competitive zero-shot tracking on four real and synthetic benchmarks. A regression readout that bypasses the video VAE and uses additional training data brings further improvement, demonstrating the potential of this direction for dense 3D tracking. We hope our analysis will inspire future work that leverages VDM priors for dense 3D tracking and 4D reconstruction. Project page: https://nirvanalan.github.io/projects/track4d
Figures & tables
Figure 1: Track4D formulates dense 3D tracking as conditional video generation. From a monocular video and point maps of a frozen 3D reconstructor, a LoRA-adapted video diffusion model generates a Residual 3D Tracking Video (metric 3D offsets) and a Normalized Coordinate Map Video (canonical coordinates), from which dense 3D tracks are recovered without per-scene optimization.
Figure 2: Track4D overview. (a) Generative readout. A frozen reconstructor supplies point maps P ; the video VAE encodes them with video V into condition C . The LoRA-adapted video DiT denoises the RTV and NCMV latents, jointly by default, and the VAE decodes the videos. Hatched targets are noised in training and start from Gaussian noise at inference. Adding inverse-normalized RTV offsets to the query-frame point map gives 3D tracks; NCMV requires 2D tracking and lifting. (b) Regression readout. Point maps enter through a patch encoder, a private motion expert attends jointly with the video DiT, and a DPT head reads residuals without the output VAE ( Sections 3.4 and 8 ). Flames denote trained modules, snowflakes frozen modules.
Figure 3: Prediction formulations (adapted from V-DPM [ 22 ] ) . Each entry Pi(tj) (short for Pi(tj,π0) , with the reference frame π0 fixed) gives the positions at time tj of the surface points observed in frame Ii . P (diagonal): geometry at each frame’s capture time. Q (column): time-invariant prediction [ 22 ] , referring every frame Ii back to the reference time t0 . R (row): dense 3D tracking, propagating the query-frame point map P0 across all timestamps. Track4D encodes R explicitly as 3D offsets ( RTV ) and Q as normalized canonical colors ( NCMV ), from which R is recovered by 2D tracking.
Figure 4: Two video representations of dense 3D tracking on DAVIS [ 70 ] . Rows show RGB, RTV offsets, NCMV canonical coordinates, and recovered XYZ tracks. Adding inverse-normalized RTV offsets to the query-frame point map recovers the tracks. All predictions are Track4D outputs.
Figure 5: Qualitative comparison of dynamic tracking on DAVIS [ 70 ] . Dense 3D tracking is visualized as colored point clouds, with color encoding query-frame identity of each tracked point. Left: input frames. Right: reconstructions from St4RTrack [ 19 ] , TraceAnything [ 31 ] , Any4D [ 23 ] , and Track4D (ours). Track4D produces more temporally coherent trajectories with fewer outliers, particularly on articulated and deformable motion ( e.g . moving mug, walking cows). All baselines are run from their official released checkpoints and inference scripts at the resolution recommended by the authors; sequences shown were selected to expose dynamic-object failure modes for the comparators. More DAVIS results appear in Figure 14 .
PointOdyssey [ 44 ]
ADT [ 89 ]
PStudio [ 90 ]
Dynamic Replica [ 45 ]
All
Dyn
All
Dyn
Dyn
All
Dyn
Method
APD ↑
EPE ↓
APD ↑
EPE ↓
APD ↑
EPE ↓
APD ↑
EPE ↓
APD ↑
EPE ↓
APD ↑
EPE ↓
APD ↑
EPE ↓
CoTracker ∗ [ 75 ]
0.6165
0.4721
0.4722
0.5264
0.8303
0.1664
0.9157
0.0397
0.7579
0.2031
0.6319
0.3842
0.6456
0.3014
SpaTrackerV2 [ 17 ]
0.2655
1.1016
0.3512
0.6921
0.5594
0.5158
0.8509
0.0621
0.5325
0.4211
0.6264
0.4052
0.5380
0.3910
St4RTrack [ 19 ]
0.6751
0.3126
0.7380
0.2146
0.7848
0.1990
0.8864
0.0526
0.8002
0.1709
0.7353
0.2651
0.7944
0.1681
TraceAnything [ 31 ]
0.2629
1.1003
0.4323
0.5798
0.8156
0.1773
0.8708
0.0559
0.6469
0.3085
0.3766
0.7415
0.5418
0.3972
Table 1: Sparse 3D point tracking ( Sim(3) aligned). APD ↑ and EPE ↓ after a global Sim(3) alignment to the ground truth, as in 4RC [ 34 ] , on all points (All) and dynamic points (Dyn). Bold and underline mark best and second-best results. CoTracker ∗ : CoTracker lifted by VGGT. Track4D -NCMV and Track4D -RTV are the two readouts of the joint Wan2.1-I2V-14B model; Track4D -NCMV predicts dynamic points only. ‡ Official checkpoint replayed with STream3R geometry under the same scorer (dynamic ADT not reported). † Track4D -Reg ( Section 3.4 ): Wan2.2-5B, STream3R scaffold, eight training sources including PointOdyssey and Dynamic Replica.
Variant
ADT [ 89 ]
PStudio [ 90 ]
DR [ 45 ]
All
Dyn
Dyn
All
Dyn
APD ↑
EPE ↓
APD ↑
EPE ↓
APD ↑
EPE ↓
APD ↑
EPE ↓
APD ↑
EPE ↓
Track4D (Wan 2.2 5B)
0.8831
0.1196
0.9370
0.0722
0.7281
0.2231
0.5245
0.4468
0.7009
0.2536
w/o Residual Pred
0.6625
0.3235
0.9218
0.0859
0.6100
0.3358
0.2959
0.8439
0.4837
0.4756
w/o Log Norm
0.8793
0.1217
0.9200
0.0837
0.7109
0.2384
0.5109
0.4668
0.6857
0.2690
w/o XYZ Cond
0.8722
0.1289
0.8974
0.1046
0.6396
0.3106
0.4386
0.5718
0.5908
0.3608
Table 2: Ablations of design choices. Baseline is Track4D (Wan 2.2 5B, STream3R). Top three rows modify one component each; bottom rows scale the VDM backbone (1.3B / 14B) and swap conditioning (DAv3 / STream3R). Green/red show gains/losses over baseline; intensity indicates magnitude. All models are trained and evaluated at 256×256 following [ 24 ] . Reported on ADT [ 89 ] , PStudio [ 90 ] , and DR [ 45 ] ; PO [ 44 ] is excluded due to mismatched aspect ratios.
Figure 6: TrackCraft3R vs . the three Track4D readouts on DAVIS [ 70 ] . Each panel shows the point cloud of frame j and the last 12 frames of the predicted 3D tracks, from one fixed view (the frame- 0 input camera rotated by 4∘ ). All methods receive the same VGGT- Ω point maps and draw the same 767 queries, the most-moving points of a dense grid, colored by image row. RTV and NCMV come from the joint Wan2.1-I2V-14B model ( 50 steps, no input-latent noise); NCMV tracks are read with CoTracker3, lifted with the same point maps, and frozen where NCMV support is lost (on camel , after the first frames). No ground truth or Sim(3) is used. More clips in Figures 9 and 10 .
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Submitted version
This version
Readouts
RTV and NCMV , generated jointly by a LoRA-adapted Wan2.1-I2V-14B
Unchanged; adds Track4D -Reg, a regression readout of the RTV residual without the output VAE ( Section 3.4 )
Generative model
Wan2.1-I2V-14B, LoRA rank 64 , point-map scaffold, Kubric and TartanAir
Unchanged checkpoint
Regression model
–
Wan2.2-TI2V-5B with a motion expert and a DPT head, eight training sources ( Section G.6 )
Main table
Four sparse benchmarks and dense Kubric
Four sparse benchmarks; adds Track4D -Reg and TrackCraft3R; dense Kubric in Table 5
Analyses
Ablations, sampling steps
Adds runtime, role of NCMV , VAE-free readout, artifact sources, camera-following audit, geometry/motion oracle ( Appendix G )
Figures
Overview of the generative model ( Figure 7 )
Overview redrawn with the regression readout ( Figure 2 ); DAVIS comparison with TrackCraft3R ( Figures 6 and 9 )
Appendix
Table 3: Submitted version versus this version. The submitted column summarizes the described design; the training run that produced the submitted numbers is distinguished below.
Figure 7: Submitted overview (reproduced from the submitted version). A monocular video V and point maps P from an off-the-shelf reconstructor are encoded by the video VAE into the condition C ; the VDM jointly generates the RTV Δ and the NCMV M , and the decoded residuals are added to the query-frame point map. Figure 2 shows the current overview.
Steps
EPE ↓
Time / Seq. ↓
Kubric [ 30 ]
PO [ 44 ]
ADT [ 89 ]
PStudio [ 90 ]
DR [ 45 ]
50
1.0463
0.3953
0.1149
0.2646
0.2549
14.0 ± 1.3 s
10
1.0394
0.3821
0.1066
0.2642
0.2548
9.6 ± 1.1 s
5
1.0862
0.3773
0.1039
0.2668
0.2610
8.6 ± 1.1 s
Appendix
Table 4: Sampling-step ablation. We ablate inference-time sampling steps using a Wan2.2-5B Track4D model trained at 256×256 on 512 Kubric scenes without point-map conditioning. Performance remains largely stable when reducing the number of sampling steps from 50 to 10, with 5 steps still competitive while further reducing runtime.
Method
All
Dyn
APD ↑
EPE ↓
APD ↑
EPE ↓
St4RTrack [ 19 ]
0.0468
5.3331
0.0546
2.8485
TraceAnything [ 31 ]
0.2879
1.5501
0.1379
1.9263
Any4D [ 23 ]
0.2242
1.9223
0.1370
1.8107
V-DPM [ 22 ]
0.1729
4.7054
0.4822
0.6845
[0pt][0pt] Track4D -RTV
0.3359
1.0790
0.1890
1.4042
Appendix
Table 5: Dense 3D tracking on the held-out Kubric split. APD ↑ / EPE ↓ after Sim(3) alignment on 50 held-out sequences of 49 frames, dense annotations; “All” and “Dyn” as in Table 1 . Kubric matches the training distribution of Track4D -RTV and is not part of the benchmark suite of Table 1 .
Method
Hardware
Runtime (s/seq.) ↓
Peak memory (GB, reserved) ↓
St4RTrack ( 5122 )
A6000 48 GB
6.07±0.03
9.56
TraceAnything ( 5122 )
A6000 48 GB
6.76±0.05
37.17
CoTracker + VGGT ( 4762 )
A6000 48 GB
7.53±0.15
23.04
4RC ( 504×378 )
A6000 48 GB
8.76±0.09
20.60
Any4D ( 5182 )
A6000 48 GB
10.07±0.14
43.31
SpatialTrackerV2 ( 5182 )
A6000 48 GB
28.00±2.71
26.29
Appendix
Table 6: Inference time and peak memory for 49 -frame sequences on one GPU. Native inference resolutions are shown in parentheses. The Track4D row measures the 5 B, 5 -step configuration, which is not the configuration used for the accuracy results in Table 1 . V-DPM does not fit on the 48 GB A6000; its H100 measurement is included only as a hardware-unmatched reference and should not be read as a direct runtime comparison. † Measured on an A100 over 90 warm DAVIS clips (five passes, BF16), excluding the production of the input point maps; also hardware-unmatched.
Benchmark
RTV-only
Joint RTV+NCMV
PointOdyssey
0.2891 / 0.6908
0.2856 / 0.6931
ADT
0.0715 / 0.9390
0.0666 / 0.9466
Panoptic Studio
0.1610 / 0.8093
0.1511 / 0.8240
Dynamic Replica
0.1886 / 0.7813
0.1861 / 0.7850
Mean (dynamic)
0.1776 / 0.8051
0.1724 / 0.8122
Mean (all-point)
0.1848 / 0.8134
0.1804 / 0.8195
Appendix
Table 7: RTV-only and joint RTV+NCMV checkpoints. Per-benchmark rows report dynamic EPE ↓ / APD ↑ on 49 -frame sequences; the final row reports all-point means. Both are Wan-14B LoRA checkpoints with different training histories, and they are not sampled identically. The joint checkpoint is the model of Table 1 : here it is sampled with 1 denoising step and CFG 1 at a 256 -pixel short side, whereas Table 1 reports its submitted evaluation with 50 denoising steps and random input-latent noise ( Section 4.1 ). The two tables therefore give different values for the same checkpoint ( e.g . PStudio dynamic EPE 0.1511 here vs . 0.1596 in Table 1 ). Bold denotes the better result.
Model
PointOdyssey
ADT
PStudio
DynamicReplica
Mean
Submitted 5B (VAE decode)
0.379 / 0.594
0.072 / 0.937
0.223 / 0.728
0.254 / 0.701
0.232 / 0.740
Direct readout, Kubric tracking
0.321 / 0.655
0.117 / 0.871
0.251 / 0.695
0.210 / 0.753
0.225 / 0.743
+ depth and camera supervision
0.317 / 0.657
0.107 / 0.883
0.216 / 0.734
0.206 / 0.758
0.212 / 0.758
+ PointOdyssey tracking
0.290 / 0.689
0.090 / 0.913
0.176 / 0.786
0.187 / 0.785
0.186 / 0.793
Appendix
Table 8: VAE-free direct readout with a Wan2.2-5B backbone. Dynamic EPE ↓ / APD ↑ ; the mean averages the four benchmarks. The first two rows compare the output paths under the same Kubric tracking data. The final row additionally uses PointOdyssey for training, so PointOdyssey is in-domain for that variant.
Figure 8: Tensor flow of the regression readout ( Track4D -Reg). Top: frame-wise RGB latents and point-map features form world tokens; repeating the query-frame features forms motion tokens. Middle: the LoRA-adapted Wan expert and private motion expert share attention and 3D rotary positions in each block. Bottom: motion features from four blocks feed a DPT head for residuals and visibility. Residuals are added to the supplied query point map; five frame-0-anchored passes cover 49 frames. Flames: trained modules; VAE and Wan base weights: frozen.
Geometry
Method
PStudio
PointOdyssey
Dynamic Replica
ADT
own
V-DPM [ 22 ]
0.0883 / 0.9224
0.1403 / 0.8642
0.1914 / 0.8184
0.1045 / 0.9024
4RC [ 34 ]
0.1067 / 0.8936
0.1939 / 0.8347
0.1604 / 0.8664
0.1144 / 0.8894
TrackCraft3R [ 66 ]
0.1017 / 0.9031
0.4370 / 0.6524
0.2239 / 0.7783
0.0935 / 0.9245
[0pt][0pt]
Track4D -Reg (ours)
0.0947 / 0.9155
0.4127 / 0.6668
0.2257 / 0.7777
0.0933 / 0.9230
VGGT- Ω
Track4D -Reg + GT query geometry ∘
0.0900 / 0.9169
0.0531 / 0.9450
0.0275 / 0.9743
0.0374 / 0.9582
TrackCraft3R [ 66 ]
0.1347 / 0.8491
0.5224 / 0.5064
0.2859 / 0.7134
0.4002 / 0.6187
Appendix
Table 9: Regression readout vs . TrackCraft3R. All-point EPE ↓ / APD ↑ on 50 sequences per benchmark ( 49 for ADT), using the protocol of Table 1 . TrackCraft3R and Track4D -Reg share point-map caches; bold marks the better result. V-DPM and 4RC use their own geometry. PointOdyssey and Dynamic Replica are training datasets for Track4D -Reg, V-DPM and 4RC. ∘ Oracle diagnostic without retraining: the ground-truth query point replaces the estimated one after prediction, scored under the fixed Sim(3) of the row above ( Table 11 ), following TrackCraft3R’s input-geometry ablation.
Figure 9: More DAVIS clips, TrackCraft3R vs . the three Track4D readouts. Same setup as Figure 6 . Because the view is fixed at frame 0 , subjects that the camera follows leave the view in later frames ( car-shadow , kite-surf , parkour ). Over all 91 DAVIS clips, the median second difference of the projected tracks over consecutive frames, counting points inside the frame, is 6.5 px for TrackCraft3R and 1.8 px for Track4D -Reg ( 672 -px-wide frames), and it is smaller for Track4D -Reg on 90 clips. This measures temporal smoothness, not accuracy.
Figure 10: More DAVIS clips (continued). Same setup as Figures 6 and 9 .
Benchmark
Wan2.2-5B ( RTV )
Wan2.1-14B joint ( RTV )
Track4D -Reg
Track4D -Reg, all points
PStudio
0.2231 / 0.7281
0.1511 / 0.8240
0.1311 / 0.8572
0.1311 / 0.8572
PointOdyssey
0.3792 / 0.5939
0.2856 / 0.6931
0.2691 / 0.7155
0.2078 / 0.8050
Dynamic Replica
0.2536 / 0.7009
0.1861 / 0.7850
0.1674 / 0.8141
0.2284 / 0.7871
ADT
0.0722 / 0.9370
0.0666 / 0.9466
0.0675 / 0.9430
0.1168 / 0.8854
Appendix
Table 10: Track4D -Reg against the generative models with the same STream3R scaffold. Dynamic EPE ↓ / APD ↑ under the WorldTrack scorer ( 50 sequences, 49 frames; ADT 49 ); the last column reports all-point values of Track4D -Reg. Bold marks the best dynamic value. See the text for the differences in raster and training data.
Fixed Sim(3)
Refit Sim(3)
Geometry
Benchmark
Current
GT query geometry
GT motion
GT query geometry
GT motion
VGGT- Ω
PStudio
0.0947 / 0.9155
0.0900 / 0.9169
0.0900 / 0.9239
0.0763 / 0.9415
0.0672 / 0.9492
PointOdyssey
0.4127 / 0.6668
0.0531 / 0.9450
0.4035 / 0.6787
0.0567 / 0.9518
0.4017 / 0.6810
Dynamic Replica
0.2257 / 0.7777
0.0275 / 0.9743
0.2207 / 0.7850
0.0344 / 0.9764
0.2202 / 0.7851
ADT (secondary)
0.0933 / 0.9230
0.0374 / 0.9582
0.1056 / 0.9105
0.0261 / 0.9745
0.0789 / 0.9459
DAv3-L
PStudio
0.1270 / 0.8625
0.1146 / 0.8761
0.1289 / 0.8614
0.0933 / 0.9147
0.1011 / 0.9050
Appendix
Table 11: Full geometry/motion oracle for Track4D -Reg. EPE ↓ / APD ↑ on all 50 sequences of each primary benchmark (ADT, 49 sequences, is secondary). Fixed columns score the substituted output under the original all-track Sim(3) of the current prediction; refit columns fit a new closed-form Sim(3) to each substituted output. Closed-form Sim(3) minimizes squared distances rather than EPE or APD, so refitting can raise EPE; neither column gives a lower bound.
Geometry
Benchmark
Total
Anchor
Motion
Cross
VGGT- Ω
PStudio
0.0219
0.0197
0.0200
−0.0178
PointOdyssey
0.6349
0.6299
0.0127
−0.0076
Dynamic Replica
0.2103
0.2074
0.0060
−0.0031
DAv3-L
PStudio
0.0307
0.0335
0.0280
−0.0308
PointOdyssey
0.3784
0.3658
0.0422
−0.0297
Dynamic Replica
0.2560
0.2558
0.0090
−0.0088
Appendix
Table 12: Exact squared-error budget of Track4D -Reg. Sequence-macro mean squared error in m 2 under the original Sim(3) , decomposed as in Equation 12 ; the identity holds to floating-point precision. The cross term is large and negative on PStudio, so anchor and motion errors must not be read as additive EPE contributions.
Checkpoint
Geometry
PStudio
PointOdyssey
Dynamic Replica
ADT
parent run, step 13,750
VGGT- Ω
0.2317 / 0.7197
0.4337 / 0.6426
0.2436 / 0.7550
0.1030 / 0.9106
STream3R
0.2420 / 0.7082
0.2350 / 0.7693
0.2476 / 0.7611
0.1234 / 0.8764
recovery stage, step 7,500
VGGT- Ω
0.0988 / 0.9080
0.4146 / 0.6648
0.2272 / 0.7758
0.0995 / 0.9157
STream3R
0.1338 / 0.8527
0.2103 / 0.8016
0.2300 / 0.7848
0.1183 / 0.8841
recovery stage, step 1,000 (reported)
VGGT- Ω
0.0947 / 0.9155
0.4127 / 0.6668
0.2257 / 0.7777
0.0933 / 0.9230
STream3R
0.1311 / 0.8572
0.2078 / 0.8050
0.2284 / 0.7871
0.1168 / 0.8854
Appendix
Table 13: Backend swap across checkpoints of the Track4D -Reg lineage. EPE ↓ / APD ↑ on all points using VGGT- Ω or STream3R point maps, with fixed weights within each checkpoint. The reported checkpoint follows Table 9 ; inference settings for the earlier checkpoints have not been fully verified, so cross-checkpoint differences are descriptive. Across checkpoints, STream3R roughly halves PointOdyssey EPE; Dynamic Replica changes little, while PStudio and ADT prefer VGGT- Ω .
Diagnostic arm
Isolates
All EPE ↓
Predicted point-map scaffold
scaffold error
≈2.34
Oracle scaffold + predicted RTV residual
residual-prediction error
≈0.84
Ground-truth target → VAE round-trip
VAE codec floor
≈0.79
Full pipeline
all stages combined
≈2.12
Appendix
Table 14: Controlled stage-wise diagnostics of geometric error. Metric-space EPE on a fixed Kubric val10 subset. The arms isolate different sources and are not additive.
Output path
All EPE ↓
Edge EPE ↓
Interior EPE ↓
VAE-decoded
1.340
1.479
1.300
VAE-free direct readout
0.859
1.331
0.728
Relative reduction
35.9%
10.0%
44.0%
Appendix
Table 15: Metric impact of removing the VAE output path. Metric-space EPE on the fixed Kubric val10 subset.
Figure 11: Failure case. Track4D predicts near-zero deformation in this video, where the camera moves with the object. Limited coverage of such camera-following-object trajectories in the training data is a plausible, but not conclusively established, cause ( Section G.9 ).
Figure 12: Ablation: absolute XYZ prediction vs. residual prediction. We visualize 3D tracking results at t=0 and t=24 on ADT [ 89 ] and PStudio [ 90 ] sequences. Left: predicting absolute 3D coordinates directly ( w/o Residual Pred ). Right: our residual formulation ( Equation 4 ). Absolute prediction produces noisy geometry with scattered outliers, whereas the residual formulation yields cleaner reconstructions with more coherent motion.
Figure 13: Effect of backbone size. Specifically, we show our model’s performance on four samples from PStudio [ 90 ] , with sequence names on the left. For each sequence, we compare the results from Track4D -5B (in orange ) and Track4D -14B (in green ). Both RTV and NCMV are visualized. The Wan-14B-based Track4D shows better foreground-background factorization and sharper motion, suggesting benefits from stronger video diffusion models.
Figure 14: More qualitative results. We show seven zero-shot inference results from DAVIS [ 70 ] . The first row shows RGB input, with eight frames sampled uniformly from a 49 -frame clip, and the second and third rows show RTV and NCMV predictions. The query frame is the first frame shown in the first row. RTV encodes 3D offsets on the query-frame grid; NCMV encodes persistent coordinates on each video frame’s grid.
Asset
Use
License / terms
Kubric [ 30 ]
Rendering our training split
Apache-2.0 (code)
TartanAir [ 88 ]
Training (generative)
CC BY 4.0
PointOdyssey [ 44 ]
Training (regression), evaluation
CC BY-NC-SA 4.0 (data); MIT (code)
Dynamic Replica [ 45 ]
Training (regression), evaluation
CC BY-NC (repository); data after accepting the license agreement
Waymo Open Dataset [ 100 ]
Training (regression)
Waymo Open Dataset terms; Apache-2.0 (code)
DL3DV-10K [ 101 ]
Training (regression)
DL3DV-10K terms of use and CC BY-NC 4.0
Appendix
Table 16: Existing assets and their licenses , as stated by the official repositories and dataset pages (checked October 2026). Data and model weights may carry terms beyond the code license; we use every asset for non-commercial research only.