Image Classification on SUN397 (Top-1 Accuracy, Delta)
80.3Top-1 AccuracyWUDI-Merging
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| WUDI-MergingMerging Method=WUDI-Merging, Quantization Method=None, Bit-width=Full-precision, Model Backbone=CLIP-ViT-L/142026.05 | 80.3 | — | |
| WUDI-Merging w/ E-PMQMerging Method=WUDI-Merging, Quantization Method=E-PMQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 79.9 | — | |
| WUDI-Merging w/ GPTQMerging Method=WUDI-Merging, Quantization Method=GPTQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 79.1 | — | |
| WUDI-Merging w/ RTNMerging Method=WUDI-Merging, Quantization Method=RTN, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 78.9 | — | |
| SDS-LoRAModel=ViT-Large, Rank=82026.06 | 76.58 | — | |
| DoRAModel=ViT-Large, Rank=82026.06 | 76.5 | — | |
| SDS-LoRAModel=ViT-Large, Rank=322026.06 | 76.46 | — | |
| rsLoRAModel=ViT-Large, Rank=82026.06 | 76.45 | — | |
| LoRAModel=ViT-Large, Rank=82026.06 | 76.3 | — | |
| DoRAModel=ViT-Large, Rank=322026.06 | 76.23 | — | |
| rsLoRAModel=ViT-Large, Rank=322026.06 | 76.22 | — | |
| Simple Averaging w/ E-PMQMerging Method=Simple Averaging, Quantization Method=E-PMQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 76.2 | — | |
| Full FT.Model=ViT-Large, Rank=-2026.06 | 76.2 | — | |
| TIES-Merging w/ E-PMQMerging Method=TIES-Merging, Quantization Method=E-PMQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 76 | — | |
| Task Arithmetic w/ E-PMQMerging Method=Task Arithmetic, Quantization Method=E-PMQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 75.9 | — | |
| PiSSAModel=ViT-Large, Rank=82026.06 | 75.41 | — | |
| LoRA+Model=ViT-Large, Rank=322026.06 | 75.19 | — | |
| LoRA+Model=ViT-Large, Rank=82026.06 | 75.08 | — | |
| Fine-tuned (STL)Backbone=CLIP-ViT-B/32, Bit-width=Full-precision, Training=Single-task learning2026.05 | 75 | — | |
| SDS-LoRARank=32, Backbone=ViT-Base2026.06 | 74.89 | — | |
| SDS-LoRAModel=ViT-Base, Rank=322026.06 | 74.89 | — | |
| SDS-LoRARank=8, Backbone=ViT-Base2026.06 | 74.85 | — | |
| SDS-LoRAModel=ViT-Base, Rank=82026.06 | 74.85 | — | |
| Full FT.Rank=-, Backbone=ViT-Base2026.06 | 74.76 | — | |
| Full FT.Model=ViT-Base, Rank=-2026.06 | 74.76 | — | |
| TIES-MergingMerging Method=TIES-Merging, Quantization Method=None, Bit-width=Full-precision, Model Backbone=CLIP-ViT-L/142026.05 | 74.7 | — | |
| rsLoRARank=8, Backbone=ViT-Base2026.06 | 74.63 | — | |
| rsLoRAModel=ViT-Base, Rank=82026.06 | 74.63 | — | |
| LoRAModel=ViT-Large, Rank=322026.06 | 74.59 | — | |
| DoRARank=8, Backbone=ViT-Base2026.06 | 74.46 | — | |
| DoRAModel=ViT-Base, Rank=82026.06 | 74.46 | — | |
| LoRARank=8, Backbone=ViT-Base2026.06 | 74.35 | — | |
| LoRAModel=ViT-Base, Rank=82026.06 | 74.35 | — | |
| BiCLIPShots=16, Base Model=CLIP, Adaptation Method=Bilinear adaptation2026.03 | 74.27 | 10.77 | |
| BiSigLIPShots=16, Base Model=SigLIP, Adaptation Method=Bilinear adaptation2026.03 | 74.24 | 8.38 | |
| WUDI-Merging w/ AWQMerging Method=WUDI-Merging, Quantization Method=AWQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 74.1 | — | |
| DoRARank=32, Backbone=ViT-Base2026.06 | 74.05 | — | |
| DoRAModel=ViT-Base, Rank=322026.06 | 74.05 | — | |
| TIES-Merging w/ GPTQMerging Method=TIES-Merging, Quantization Method=GPTQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 74 | — | |
| rsLoRARank=32, Backbone=ViT-Base2026.06 | 73.92 | — | |
| rsLoRAModel=ViT-Base, Rank=322026.06 | 73.92 | — | |
| PiSSAModel=ViT-Large, Rank=322026.06 | 73.84 | — | |
| TIES-Merging w/ RTNMerging Method=TIES-Merging, Quantization Method=RTN, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 73.8 | — | |
| LoRA+Rank=8, Backbone=ViT-Base2026.06 | 72.87 | — | |
| LoRA+Model=ViT-Base, Rank=82026.06 | 72.87 | — | |
| PiSSARank=8, Backbone=ViT-Base2026.06 | 72.82 | — | |
| PiSSAModel=ViT-Base, Rank=82026.06 | 72.82 | — | |
| Simple AveragingMerging Method=Simple Averaging, Quantization Method=None, Bit-width=Full-precision, Model Backbone=CLIP-ViT-L/142026.05 | 72.5 | — | |
| Simple Averaging w/ GPTQMerging Method=Simple Averaging, Quantization Method=GPTQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 72 | — | |
| Task ArithmeticMerging Method=Task Arithmetic, Quantization Method=None, Bit-width=Full-precision, Model Backbone=CLIP-ViT-L/142026.05 | 72 | — | |
| Simple Averaging w/ RTNMerging Method=Simple Averaging, Quantization Method=RTN, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 71.6 | — | |
| Task Arithmetic w/ GPTQMerging Method=Task Arithmetic, Quantization Method=GPTQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 71.3 | — | |
| Task Arithmetic w/ RTNMerging Method=Task Arithmetic, Quantization Method=RTN, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 71 | — | |
| LoRARank=32, Backbone=ViT-Base2026.06 | 70.82 | — | |
| LoRAModel=ViT-Base, Rank=322026.06 | 70.82 | — | |
| TIES-Merging w/ AWQMerging Method=TIES-Merging, Quantization Method=AWQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 70.4 | — | |
| LR-LoRABackbone=DINOv2 ViT-B/14, Rank=16, Adapter Placement=attention and MLP projections, Evaluation Seeds=32026.06 | 70.2 | — | |
| PiSSARank=32, Backbone=ViT-Base2026.06 | 69.73 | — | |
| PiSSAModel=ViT-Base, Rank=322026.06 | 69.73 | — | |
| RandLoRABackbone=DINOv2 ViT-B/14, Rank=16, Adapter Placement=attention and MLP projections, Evaluation Seeds=32026.06 | 69.3 | — | |
| LoRABackbone=DINOv2 ViT-B/14, Rank=16, Adapter Placement=attention and MLP projections, Evaluation Seeds=32026.06 | 68.9 | — | |
| VeRABackbone=DINOv2 ViT-B/14, Rank=16, Adapter Placement=attention and MLP projections, Evaluation Seeds=32026.06 | 68.5 | — | |
| WUDI-MergingBackbone=CLIP-ViT-B/32, Merging Strategy=WUDI-Merging, Bit-width=Full-precision2026.05 | 68 | — | |
| WUDI-Merging w/ E-PMQBackbone=CLIP-ViT-B/32, Merging Strategy=WUDI-Merging, Quantization Algorithm=E-PMQ, Bit-width=4-bit2026.05 | 67.9 | — | |
| Simple Averaging w/ E-PMQBackbone=CLIP-ViT-B/32, Merging Strategy=Simple Averaging, Quantization Algorithm=E-PMQ, Bit-width=4-bit2026.05 | 67.7 | — | |
| LoRA+Rank=32, Backbone=ViT-Base2026.06 | 67.66 | — | |
| LoRA+Model=ViT-Base, Rank=322026.06 | 67.66 | — | |
| TIES-Merging w/ E-PMQBackbone=CLIP-ViT-B/32, Merging Strategy=TIES-Merging, Quantization Algorithm=E-PMQ, Bit-width=4-bit, Lambda=0.32026.05 | 67.6 | — | |
| WUDI-Merging w/ E-PMQMerging Strategy=WUDI-Merging, Quantization Method=E-PMQ, Model Backbone=CLIP-ViT-B/32, Bit-width=4-bit2026.05 | 67.44 | — | |
| TIES-MergingBackbone=CLIP-ViT-B/32, Merging Strategy=TIES-Merging, Bit-width=Full-precision, Lambda=0.32026.05 | 67.1 | — | |
| Task Arithmetic w/ E-PMQBackbone=CLIP-ViT-B/32, Merging Strategy=Task Arithmetic, Quantization Algorithm=E-PMQ, Bit-width=4-bit, Lambda=0.32026.05 | 67 | — | |
| WUDI-Merging w/ GPTQBackbone=CLIP-ViT-B/32, Merging Strategy=WUDI-Merging, Quantization Algorithm=GPTQ, Bit-width=4-bit2026.05 | 66.6 | — | |
| Simple Averaging w/ AWQMerging Method=Simple Averaging, Quantization Method=AWQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 66.6 | — | |
| SigLIPShots=0, Base Model=SigLIP, Adaptation Method=Zero-Shot2026.03 | 65.85 | — | |
| WUDI-MergingMerging Strategy=WUDI-Merging, Quantization Method=None, Model Backbone=CLIP-ViT-B/32, Bit-width=Full2026.05 | 65.7 | — | |
| Simple Averaging w/ E-PMQMerging Strategy=Simple Averaging, Quantization Method=E-PMQ, Model Backbone=CLIP-ViT-B/32, Bit-width=4-bit2026.05 | 65.65 | — | |
| TIES-Merging w/ GPTQBackbone=CLIP-ViT-B/32, Merging Strategy=TIES-Merging, Quantization Algorithm=GPTQ, Bit-width=4-bit, Lambda=0.32026.05 | 65.6 | — | |
| Simple AveragingBackbone=CLIP-ViT-B/32, Merging Strategy=Simple Averaging, Bit-width=Full-precision2026.05 | 65.4 | — | |
| TIES-Merging w/ E-PMQMerging Strategy=TIES-Merging, Quantization Method=E-PMQ, Model Backbone=CLIP-ViT-B/32, Bit-width=4-bit2026.05 | 65.28 | — | |
| Simple AveragingMerging Strategy=Simple Averaging, Quantization Method=None, Model Backbone=CLIP-ViT-B/32, Bit-width=Full2026.05 | 64.8 | — | |
| Simple Averaging + E-PMQMerging Strategy=Simple Averaging, Quantization Method=E-PMQ, Bit-width=4-bit, Backbone=CLIP-ViT-B/322026.05 | 64.59 | — | |
| WUDI-Merging + E-PMQMerging Strategy=WUDI-Merging, Quantization Method=E-PMQ, Bit-width=4-bit, Backbone=CLIP-ViT-B/322026.05 | 64.37 | — | |
| Simple Averaging w/ GPTQBackbone=CLIP-ViT-B/32, Merging Strategy=Simple Averaging, Quantization Algorithm=GPTQ, Bit-width=4-bit2026.05 | 64.3 | — | |
| Task Arithmetic w/ E-PMQMerging Strategy=Task Arithmetic, Quantization Method=E-PMQ, Model Backbone=CLIP-ViT-B/32, Bit-width=4-bit2026.05 | 64.23 | — | |
| Simple AveragingMerging Strategy=Simple Averaging, Quantization Method=Full-precision, Bit-width=Full, Backbone=CLIP-ViT-B/322026.05 | 64.2 | — | |
| TIES-Merging + E-PMQMerging Strategy=TIES-Merging, Quantization Method=E-PMQ, Bit-width=4-bit, Backbone=CLIP-ViT-B/322026.05 | 64.04 | — | |
| Simple Averaging w/ GPTQMerging Strategy=Simple Averaging, Quantization Method=GPTQ, Model Backbone=CLIP-ViT-B/32, Bit-width=4-bit2026.05 | 63.89 | — | |
| WUDI-Merging w/ GPTQMerging Strategy=WUDI-Merging, Quantization Method=GPTQ, Model Backbone=CLIP-ViT-B/32, Bit-width=4-bit2026.05 | 63.77 | — | |
| CLIPShots=0, Base Model=CLIP, Adaptation Method=Zero-Shot2026.03 | 63.5 | — | |
| Pre-trainedBackbone=CLIP-ViT-B/32, Bit-width=Full-precision2026.05 | 63.2 | — | |
| TIES-Merging w/ RTNBackbone=CLIP-ViT-B/32, Merging Strategy=TIES-Merging, Quantization Algorithm=RTN, Bit-width=4-bit, Lambda=0.32026.05 | 63.1 | — | |
| TIES-Merging w/ AWQBackbone=CLIP-ViT-B/32, Merging Strategy=TIES-Merging, Quantization Algorithm=AWQ, Bit-width=4-bit, Lambda=0.32026.05 | 62.9 | — | |
| WUDI-Merging w/ RTNBackbone=CLIP-ViT-B/32, Merging Strategy=WUDI-Merging, Quantization Algorithm=RTN, Bit-width=4-bit2026.05 | 62.9 | — | |
| Simple Averaging + GPTQMerging Strategy=Simple Averaging, Quantization Method=GPTQ, Bit-width=4-bit, Backbone=CLIP-ViT-B/322026.05 | 62.79 | — | |
| Task Arithmetic w/ AWQMerging Method=Task Arithmetic, Quantization Method=AWQ, Bit-width=4-bit, Model Backbone=CLIP-ViT-L/142026.05 | 62.6 | — | |
| WUDI-Merging w/ AWQBackbone=CLIP-ViT-B/32, Merging Strategy=WUDI-Merging, Quantization Algorithm=AWQ, Bit-width=4-bit2026.05 | 62.5 | — | |
| TIES-MergingMerging Strategy=TIES-Merging, Quantization Method=None, Model Backbone=CLIP-ViT-B/32, Bit-width=Full2026.05 | 62.2 | — | |
| Simple Averaging w/ RTNBackbone=CLIP-ViT-B/32, Merging Strategy=Simple Averaging, Quantization Algorithm=RTN, Bit-width=4-bit2026.05 | 62 | — | |
| Task Arithmetic + E-PMQMerging Strategy=Task Arithmetic, Quantization Method=E-PMQ, Bit-width=4-bit, Backbone=CLIP-ViT-B/322026.05 | 61.83 | — | |
| WUDI-Merging w/ AWQMerging Strategy=WUDI-Merging, Quantization Method=AWQ, Model Backbone=CLIP-ViT-B/32, Bit-width=4-bit2026.05 | 61.75 | — |