Linear Probing on ImageNet 256x256
74.6AccuracySD-VAE + SiT-XL/2 + REPA
Evaluation Results
| Method | Links | |
|---|---|---|
| SD-VAE + SiT-XL/2 + REPATokenizer=SD-VAE [1], Generator=SiT-XL/2 + REPA [71] 675M, Type=Diff.2025.04 | 74.6 | |
| GigaTok-XL-XXL + LlamaGen-XXL (1d)Tokenizer rFID=0.79, Generator=LlamaGen-XXL (1d) [50] 1.4B, Type=AR2025.04 | 74 | |
| GigaTok-XL-XXL + LlamaGen-B (1d)Tokenizer rFID=0.79, Generator=LlamaGen-B (1d) [50] 111M, Type=AR2025.04 | 72 | |
| B-AE-d32 + BiGR-XL-d32Tokenizer=B-AE-d32 [22], Generator=BiGR-XL-d32 [22] 799M, Type=AR+Diff2025.04 | 69.8 | |
| GigaTok-B-L + LlamaGen-XXL (1d)Tokenizer rFID=0.81, Generator=LlamaGen-XXL (1d) [50] 1.4B, Type=AR2025.04 | 69.4 | |
| GigaTok-B-L + LlamaGen-B (1d)Tokenizer rFID=0.51‡, Generator=LlamaGen-B (1d) [50] 111M, Type=AR2025.04 | 67.7 | |
| GigaTok-B-L + LlamaGen-B (1d)Tokenizer rFID=0.81, Generator=LlamaGen-B (1d) [50] 111M, Type=AR2025.04 | 67.6 | |
| GigaTok-S-B + LlamaGen-B (1d)Tokenizer rFID=0.89, Generator=LlamaGen-B (1d) [50] 111M, Type=AR2025.04 | 62.9 | |
| GigaTok-S-S + LlamaGen-B (1d)Tokenizer rFID=1.01, Generator=LlamaGen-B (1d) [50] 111M, Type=AR2025.04 | 62.6 | |
| VAE + MAR-HTokenizer=VAE [35], Generator=MAR-H [35] 943M, Type=AR+Diff.2025.04 | 60 | |
| LlamaGen-Tok. + LlamaGen-LTokenizer=LlamaGen-Tok. [50], Generator=LlamaGen-L [50] 343M, Type=AR2025.04 | 40.5 |