Hugging Face's logo

Motif-Technologies
/
Motif-Vision-Encoder

Image Feature Extraction
Transformers
Safetensors
motif_vision
feature-extraction
motif
vision-transformer
self-supervised
video
custom_code
Model card Files
xet
 Community
5

docs(README): add architecture figure to Model details

#4
by gkalstn0 - opened
base: refs/heads/main
←
from: refs/pr/4
Discussion Files changed
+7
-23
initial commit91172923
 Initial release: MOTIF vision encoder 7B (bf16 safetensors + modeling_motif.py)cfeb2ba5
 add linear-probe classifier ckpt (ImageNet-1K, best top1=87.266) for eval-only reproduction3018ca9e
 remove linear-probe eval ckpt538bee61
 License->MIT; strip training-only code (init_weights, torch._dynamo config, stochastic depth). Inference forward bit-identical (verified).c436f1d5
 refactor: consolidate all imports to module top (dedupe per-section imports)d9820db4
 rename modeling_motif.py -> modeling_motif_vision_encoder.py (update auto_map + README)e40c8830
 docs: update repo id to Motif-Vision-Encoder-7B in usage + citatione9288a8f
 docs: brand name MOTIF -> Motif (title, intro, citation, docstrings)6050c216
 docs: add training-data scale (~0.47B samples, 96% image / 4% video) + data-efficiency vs DINOv3 LVD-1689M5516ae90
 docs(README): DAVIS S/M/L table, Publisher column, updated metricsa8fa992b
 docs(README): Training-data column + corrected ImageNet numbers2517d39b
 docs(README): unify Training data column to B unitsfe722d61
 docs(README): add attention demo GIF + data-efficiency headline46f5cd48
 docs(README): swap DINOv2 for PEcore/OpenCLIP/VideoMAEv2, drop caption block ref6a5ae7ed
 docs(README): blank OpenCLIP zero-shot cell; VideoMAEv2 data as 1.35M87cb15da
 weights: publish ckpt-199 (phase3 366999base_w_gram_6n_syn, iter 199)61469903
 docs(README): Motif K400 = 87.4 (eval complete)94df3ff9
 docs(README): V-JEPA 2.1 KITTI = 2.5c0154d54
 docs(README): underline second-best value per column744127bc
 docs(README): condense results prose to DAVIS-SOTA + competitive summary0fc183e4
 docs(README): correct demo caption to point-tracking via feature similarityaed26c95
 docs(README): note Motif > V-JEPA 2.1 in point-tracking demo6586ff01
 docs(README): fix leftover img alt text to match tracking captionf9b7d75a
 assets: high-res tracking demo GIF with legible labels (#2)a32fa83c
 Update README.mdc133d3b4
 Update README.md80186058
 Update README.md8bc3bc2f
 Update README.mdb4d78cc1
 docs(README): add Web-DINO row; fix typo + stale KITTI line; tighten prose13d47d85
 docs(README): drop KITTI depth column5907c0ce
 docs(README): remove Citation sectiondae65f14
 docs(README): VideoMAEv2 data 1.35M -> 0.0014B for unit consistencyab03ef87
 docs(README): drop VideoMAEv2 row; update repo id after rename to Motif-Vision-Encoder9e057d68
gkalstn0
 Motif Technologies org

Adds an architecture diagram (image/video inputs, patch embedding, 40-block transformer stack, and a transformer-block internals callout) at the top of the Model details section, pairing the visual with the spec table. Figure is deterministic matplotlib output at 2791px width.
docs(README): add architecture figure to Model detailscadd264c
 assets: express temporal 2x compression (T -> T/2) in the architecture figure88183663
 assets: temporal-merge cue inline on the video path (2 frames -> 1 token)68184b4a
 assets: resolve caption collisions and polish figure geometryd7ffd6b0
 assets: temporal cue caption T -> T/2314a4caa
gkalstn0 changed pull request status to merged
gkalstn0 deleted the refs/pr/4 ref

Sign up or log in to comment