List your company
ResearchFormerly WizWand

Multi-task Image Classification on 8-task vision suite

94.3Average PerformanceTALL-Masks
65.1872.7480.387.86Mar 28, 2025Jun 11, 2025Aug 26, 2025Nov 10, 2025Jan 24, 2026Apr 10, 2026Jun 25, 2026
Updated 4mo ago

Evaluation results

MethodLinks
TALL-MasksBackbone=ViT-L/14, Task Index=Required, Params.=2.2×2025.0394.3
IndividualBackbone=ViT-L/14, Task Index=-, Params.=8×2025.0394.2
TSV-CBackbone=ViT-L/14, Task Index=Required, Params.=2.5×2025.0393.8
TSV-M+AdaRankBackbone=ViT-L/14, Task Index=-, Params.=1×2025.0393.7
CART+AdaRankBackbone=ViT-L/14, Task Index=-, Params.=1×2025.0393.5
EMR-MergingBackbone=ViT-L/14, Task Index=Required, Params.=4.4×2025.0393.5
TA+AdaRankBackbone=ViT-L/14, Task Index=-, Params.=1×2025.0393
ReTeX (Sample-wise)Backbone=CLIP ViT-B/32, Task-ID=Unknown, Inference mode=Sample-wise, Inference cost (per input)=0.05s, VRAM (GB)=2.52026.0692.2
ReTeX (Group: B = 64)Backbone=CLIP ViT-B/32, Task-ID=Unknown, Inference mode=Group, Batch size (B)=64, Inference cost (per input)=0.0015s, VRAM (GB)=2.72026.0692.2
ReTeX-E (Sample-wise)Backbone=CLIP ViT-B/32, Task-ID=Unknown, Inference mode=Sample-wise, Inference cost (per input)=0.07s, VRAM (GB)=1.82026.0692
ReTeX-E (Group: B = 64)Backbone=CLIP ViT-B/32, Task-ID=Unknown, Inference mode=Group, Batch size (B)=64, Inference cost (per input)=0.0015s, VRAM (GB)=2.02026.0692
TALL-MasksBackbone=ViT-B/32, Task Index=Required, Params.=2.2×2025.0390.8
IndividualBackbone=ViT-B/32, Task Index=-, Params.=8×2025.0390.5
ID-EMR-MergingBackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Component selection, Inference cost (per input)=0.004s, VRAM (GB)=1.92026.0690.5
WEMoEBackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Dynamic, Inference cost (per input)=0.02s, VRAM (GB)=4.32026.0690.4
ID-LoRABackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Component selection, Inference cost (per input)=0.0009s, VRAM (GB)=1.92026.0689.3
CART+AdaRankBackbone=ViT-B/32, Task Index=-, Params.=1×2025.0389.2
TSV-CBackbone=ViT-B/32, Task Index=Required, Params.=2.5×2025.0389
DaWinBackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Dynamic, Inference cost (per input)=0.63s, VRAM (GB)=5.52026.0689
TSV-M+AdaRankBackbone=ViT-B/32, Task Index=-, Params.=1×2025.0388.9
EMR-MergingBackbone=ViT-B/32, Task Index=Required, Params.=4.4×2025.0388.8
MoW-MergingBackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Dynamic, Inference cost (per input)=0.008s, VRAM (GB)=4.92026.0688.1
TA+AdaRankBackbone=ViT-B/32, Task Index=-, Params.=1×2025.0387.9
Twin-MergingBackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Dynamic, Inference cost (per input)=0.03s, VRAM (GB)=3.22026.0684
TIES-MergingBackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Static, Inference cost (per input)=0.0004s, VRAM (GB)=1.32026.0675.1
Consensus TABackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Static, Inference cost (per input)=0.0004s, VRAM (GB)=1.32026.0675
Task ArithmeticBackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Static, Inference cost (per input)=0.0004s, VRAM (GB)=1.32026.0670.8
Weight AveragingBackbone=CLIP ViT-B/32, Task-ID=Unknown, Model Merging type=Static, Inference cost (per input)=0.0004s, VRAM (GB)=1.32026.0666.3