Commit History
-
docs(README): fix leftover img alt text to match tracking caption
Verified signatureUpdate <img> alt from 'Self-attention visualization' to 'Point tracking: Motif vs V-JEPA 2.1' for consistency with the caption.f9b7d75 -
docs(README): note Motif > V-JEPA 2.1 in point-tracking demo
Verified signatureCaption: GIF is a top=Motif / bottom=V-JEPA 2.1 comparison grid; add that Motif tracks the subject more stably.6586ff0 -
docs(README): correct demo caption to point-tracking via feature similarity
Verified signatureThe GIF is a point-query tracking demo (query point relayed by argmax of last-frame cosine-similarity), not generic self-attention. Fix caption accordingly.aed26c9 -
docs(README): condense results prose to DAVIS-SOTA + competitive summary
Verified signatureReplace the detailed per-benchmark bullet list with a concise summary: SOTA on DAVIS, competitive elsewhere on ~1/3 of DINOv3's data; keep KITTI-pending note.0fc183e -
docs(README): underline second-best value per column
Verified signatureAdd <u>underline</u> to the runner-up in each column (comparable values only); document convention in caption.744127b -
docs(README): V-JEPA 2.1 KITTI = 2.5
Verified signatureReplace V-JEPA 2.1 KITTI placeholder 3.x with 2.5.c0154d5 -
docs(README): Motif K400 = 87.4 (eval complete)
Verified signatureFill Motif K400 attentive-probe top-1 (87.4); only KITTI remains in progress.94df3ff -
docs(README): blank OpenCLIP zero-shot cell; VideoMAEv2 data as 1.35M
Verified signatureOpenCLIP ImageNet -> – (drop zero-shot value and † footnote); VideoMAEv2 training data 0.0014B -> 1.35M; update caption + protocol note references.87cb15d -
docs(README): swap DINOv2 for PEcore/OpenCLIP/VideoMAEv2, drop caption block ref
Verified signatureRemove DINOv2 row; add PEcore (DINOv3-report numbers), OpenCLIP (2B, zero-shot IN1k footnoted), VideoMAEv2 (1.35M clips, fine-tuned K400 footnoted); move ImageNet/K400 bold to PEcore; remove 'block 20' from GIF caption; update bullets + protocol note.6a5ae7e -
docs(README): add attention demo GIF + data-efficiency headline
Verified signatureAdd assets/haaland_full_attn_blk20.gif (LFS) self-attention visualization; track *.gif via LFS; add topline noting ~1/3 of DINOv3's data (0.5B vs 1.7B) with competitive performance.46f5cd4 -
docs(README): unify Training data column to B units
Verified signatureDINOv2 142M -> 0.14B, V-JEPA 2.1 22M -> 0.022B for consistent billion-scale units.fe722d6 -
docs(README): Training-data column + corrected ImageNet numbers
Verified signatureReplace Publisher column with Training data (Motif 0.5B, DINOv3 1.7B, DINOv2 142M, V-JEPA 22M, SigLIP2 10B); correct ImageNet-1K lin. probe (DINOv3 88.4, DINOv2 87.3, SigLIP2 89.1); move ImageNet bold to SigLIP2; update narrative bullets accordingly.2517d39 -
docs(README): DAVIS S/M/L table, Publisher column, updated metrics
Verified signatureReplace Params column with Publisher (Meta/Google/Motif); move DAVIS to front and split into S/M/L per DINOv3 Table 5; fill baseline numbers from Table 5; Motif DAVIS 74.0/80.5/83.5, ImageNet 87.2, ADE20K 52.0.a8fa992
-
docs: add training-data scale (~0.47B samples, 96% image / 4% video) + data-efficiency vs DINOv3 LVD-1689M
Verified signature5516ae9 -
docs: brand name MOTIF -> Motif (title, intro, citation, docstrings)
Verified signature6050c21 -
docs: update repo id to Motif-Vision-Encoder-7B in usage + citation
Verified signaturee9288a8 -
rename modeling_motif.py -> modeling_motif_vision_encoder.py (update auto_map + README)
Verified signaturee40c883 -
License->MIT; strip training-only code (init_weights, torch._dynamo config, stochastic depth). Inference forward bit-identical (verified).
Verified signaturec436f1d
-
Initial release: MOTIF vision encoder 7B (bf16 safetensors + modeling_motif.py)
Verified signaturecfeb2ba