V-JEPA 2 l256-embed

Click a block to read its description, or select it with Tab and Enter.

V-JEPA 2 l256-embedEmbed, 64 frames at 256 × 256. Video layout F × C × H × W, excluding batch.LibreYOLOV-JEPA 2 l256-embedEmbed, 64 frames at 256 × 256. Video layout F × C × H × W, excluding batch.Video encoderVideo clip64 × 3 × 256 × 256Permute to C, F, H, W3 × 64 × 256 × 256Conv3d tubeletsKernel/stride 2×16×16, bias=TrueFlatten spatiotemporal grid8192 × 1024; grid 32 × 16 × 16Encoder block8192 × 1024, repeats=24Final LayerNorm8192 × 1024, eps=1e-6No CLS token or absolute position embedding.3D rotary position applies to Q and K.Mean over all tokens1024L2 normalize1024embed_tokens also exposes 32 × 16 × 16 × 1024.Encoder blockInput 8192 × 1024LayerNorm8192 × 1024Multi-head self-attention16 heads+LayerNorm8192 × 1024Linear1024 to 4096GELU8192 × 4096Linear4096 to 1024+Output 8192 × 1024Encoder rotary self-attentionQuery input8192 × 1024Key/value input8192 × 1024Linear Q1024 to 1024Reshape heads16 × 8192 × 64Linear K1024 to 1024Reshape heads16 × 8192 × 64Linear V1024 to 1024Reshape heads16 × 8192 × 64Rotary positionRotate time, height, widthRotary positionRotate time, height, widthMatMul Q K-transpose16 × 8192 × 8192ScaleDivide by sqrt(64)Softmax over keys16 × 8192 × 8192MatMul attention × V16 × 8192 × 64Merge heads8192 × 1024Linear output1024 to 1024Factorized 3D rotaryQ or K head8192 × 64Split channel slicesTime 20, H 20, W 20, rest 4Time rotary slice8192 × 20Height rotary slice8192 × 20Width rotary slice8192 × 20Concat rotated slices and remainder8192 × 64Remainder channels are unchanged.V receives no rotary transformation.Primitive slice definition is below.Rotary slice primitiveInput channel slice8192 × 20Axis position IDsTime-major, then row, then columnMultiply position × frequencyomega[j] = 10000^(-2j/20)Sin and duplicate half tableConcatenate half-width table twiceCos and duplicate half tableConcatenate half-width table twicePair adjacent channels[a,b] becomes [-b,a]Multiply rotation × sin8192 × 20Multiply input × cos8192 × 20+Preserves the upstream adjacent-pair rotation with duplicated half-width sine/cosine tables.Variant valuesSizeSEnhMl256256102424164096h256256128032165120g256256140840226144g384384140840226144Encoder artifacts: F=64 for every size.Released probes:l256 ssv2: F=16, 174 classesl256 diving48: F=32, 48 classesg384 ssv2: F=64, 174 classesg384 diving48: F=32, 48 classesN=(F/2) × (S/16)². r=2 floor(floor((E/h)/3)/2).u=(E/h)-3r. For l/g: r=20,u=4; h256: r=26,u=2.E: width; n: depth; h: heads; M: encoder MLP width.Pooler MLP is 4E, including g variants.Source: libreyolo/models/vjepa2/nn.py. Revision a4d0ecc9e17f.libreyolo.com