Image Reconstruction on ImageNet-1K 1.0 (val)
0.62rFIDSD-VAE
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SD-VAETokenizer Type=Continuous Latent, Num Tokens=10242026.03 | 0.62 | 26.04 | 83.4 | 63.65 | — | — | — | — | — | — | — | — | |
| Cosmos CI-16×16Tokenizer Type=Continuous Latent, Num Tokens=2562026.03 | 0.72 | 29.1 | 67.5 | 39.48 | — | — | — | — | — | — | — | — | |
| TexTok-256Tokenizer Type=Continuous Latent, Num Tokens=2562026.03 | 0.73 | 24.45 | 66.8 | 18.18 | — | — | — | — | — | — | — | — | |
| DC-AETokenizer Type=Continuous Latent, Num Tokens=642026.03 | 0.77 | 23.93 | 76.6 | 60.66 | — | — | — | — | — | — | — | — | |
| DiToTokenizer Type=Diffusion-based, Num Tokens=2562026.03 | 0.78 | 24.1 | 70.6 | 0.19 | — | — | — | — | — | — | — | — | |
| Cosmos CI-8×8Tokenizer Type=Continuous Latent, Num Tokens=10242026.03 | 0.79 | 30.18 | 80.4 | 27.66 | — | — | — | — | — | — | — | — | |
| Cosmos DI-8×8Tokenizer Type=Vector-Quantized, Num Tokens=10242026.03 | 0.87 | 24.82 | 76.3 | 26.28 | — | — | — | — | — | — | — | — | |
| FluxAutoencoder type=Flux2026.04 | 0.87 | 31.7 | 0.9105 | — | — | — | 0.0354 | — | — | — | — | — | |
| Flux + FT-SEAutoencoder type=Flux + FT-SE2026.04 | 0.9 | 31.62 | 0.9096 | — | — | — | 0.0363 | — | — | — | — | — | |
| LDM# tokens=32x32, global=false, continuous=true2025.10 | 0.9 | — | — | — | — | — | — | — | — | — | — | — | |
| Ours (After 1st stage)Tokenizer Type=Diffusion-based, Num Tokens=128, Training Stage=1st stage2026.03 | 0.91 | 23.27 | 75.2 | 2.76 | — | — | — | — | — | — | — | — | |
| Flux + Structured state-space regularizationAutoencoder type=Flux + Ours2026.04 | 0.91 | 31.55 | 0.9085 | — | — | — | 0.0374 | — | — | — | — | — | |
| Flux + EQVAEAutoencoder type=Flux + EQVAE2026.04 | 0.94 | 31.53 | 0.9082 | — | — | — | 0.0373 | — | — | — | — | — | |
| FlowMoTokenizer Type=Diffusion-based, Num Tokens=2562026.03 | 0.95 | 22.07 | 64.9 | 1.44 | — | — | — | — | — | — | — | — | |
| Ours (After 2nd stage)Tokenizer Type=Diffusion-based, Num Tokens=128, Training Stage=2nd stage (One-step / Multi-Scale)2026.03 | 1.09 | 24.74 | 80 | 87.16 | — | — | — | — | — | — | — | — | |
| OpenMagViT-V2Tokenizer Type=Vector-Quantized, Num Tokens=2562026.03 | 1.17 | 21.63 | 64 | 42.91 | — | — | — | — | — | — | — | — | |
| ViT-VQGANTokenizer Type=Vector-Quantized, Num Tokens=10242026.03 | 1.28 | — | — | 94.96 | — | — | — | — | — | — | — | — | |
| FlexTok d18-d28# tokens=1-256, global=true, continuous=false2025.10 | 1.45 | — | — | — | — | — | — | — | — | — | — | — | |
| FlexTok d18-d18# tokens=1-256, global=true, continuous=false2025.10 | 1.61 | — | — | — | — | — | — | — | — | — | — | — | |
| TiTok-B# tokens=64, global=true, continuous=false2025.10 | 1.7 | — | — | — | — | — | — | — | — | — | — | — | |
| TiTok-S-128Tokenizer Type=Vector-Quantized, Num Tokens=1282026.03 | 1.71 | 17.52 | 43.7 | 7.31 | — | — | — | — | — | — | — | — | |
| TiTok-S# tokens=128, global=true, continuous=false2025.10 | 1.71 | — | — | — | — | — | — | — | — | — | — | — | |
| RepTok# tokens=1, global=true, continuous=true2025.10 | 1.85 | — | — | — | — | — | — | — | — | — | — | — | |
| VQGAN#Params=72M, #Tokens=2562026.01 | 2.19 | — | — | — | — | — | — | — | — | — | — | — | |
| LlamaGen-16Tokenizer Type=Vector-Quantized, Num Tokens=2562026.03 | 2.19 | 20.67 | 58.9 | 4.55 | — | — | — | — | — | — | — | — | |
| LlamaGen# tokens=16x16, global=true, continuous=false2025.10 | 2.19 | — | — | — | — | — | — | — | — | — | — | — | |
| TiTok-L-32#Params=641M, #Tokens=322026.01 | 2.21 | — | — | — | — | — | — | — | — | — | — | — | |
| TiTok-L# tokens=32, global=true, continuous=false2025.10 | 2.21 | — | — | — | — | — | — | — | — | — | — | — | |
| NativeTok32#Params=616M, #Tokens=322026.01 | 2.57 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosmos + Structured state-space regularizationAutoencoder type=Cosmos + Ours2026.04 | 2.82 | 25.18 | 0.7639 | — | — | — | 0.1404 | — | — | — | — | — | |
| Cosmos + EQVAEAutoencoder type=Cosmos + EQVAE2026.04 | 2.85 | 25.2 | 0.7632 | — | — | — | 0.137 | — | — | — | — | — | |
| Cosmos + FT-SEAutoencoder type=Cosmos + FT-SE2026.04 | 2.87 | 25.37 | 0.7673 | — | — | — | 0.1335 | — | — | — | — | — | |
| CosmosAutoencoder type=Cosmos2026.04 | 2.88 | 25.16 | 0.7706 | — | — | — | 0.1353 | — | — | — | — | — | |
| ChannelTokEncoder Parameters (M)=65, Decoder Parameters (M)=95, Total Parameters (M)=159, Token Budget=5122026.06 | 2.92 | — | — | — | — | — | 0.153 | 9.6 | 51.4 | 41.3 | 19.5 | 24.2 | |
| FlexTokEncoder Parameters (M)=85, Decoder Parameters (M)=172, Total Parameters (M)=341, Token Budget=2562026.06 | 2.97 | — | — | — | — | — | 0.228 | 15.4 | 61.33 | 4.78 | 16.3 | 209.3 | |
| ChannelTokEncoder Parameters (M)=65, Decoder Parameters (M)=95, Total Parameters (M)=159, Token Budget=2562026.06 | 3.7 | — | — | — | — | — | 0.169 | 11.1 | 51.34 | 41.39 | 19.5 | 24.2 | |
| FlexTok d12-d12# tokens=1-256, global=true, continuous=false2025.10 | 4.2 | — | — | — | — | — | — | — | — | — | — | — | |
| Cosmos DI-16×16Tokenizer Type=Vector-Quantized, Num Tokens=2562026.03 | 4.4 | 19.98 | 53.6 | 9.55 | — | — | — | — | — | — | — | — | |
| KARLEncoder Parameters (M)=101, Decoder Parameters (M)=105, Total Parameters (M)=239, Token Budget=2562026.06 | 5.74 | — | — | — | — | — | 0.154 | 11.6 | 63.54 | 54.66 | 15.7 | 18.3 | |
| OneDPieceEncoder Parameters (M)=304, Decoder Parameters (M)=307, Total Parameters (M)=642, Token Budget=2562026.06 | 7.61 | — | — | — | — | — | 0.18 | 10.9 | 32.84 | 24.77 | 30.5 | 40.4 | |
| ALITEncoder Parameters (M)=202, Decoder Parameters (M)=206, Total Parameters (M)=431, Token Budget=2562026.06 | 9.74 | — | — | — | — | — | 0.18 | 11 | 18.51 | 18.34 | 54.01 | 54.54 | |
| DOVEEncoder Parameters (M)=70, Decoder Parameters (M)=70, Total Parameters (M)=287, Token Budget=2562026.06 | 13.82 | — | — | — | — | — | 0.153 | 13.3 | 118.7 | 68.83 | 8.4 | 14.5 | |
| BaselineBackbone=ConvNeXtV2-Base2025.09 | — | — | — | — | 0.3151 | 0.2415 | — | — | — | — | — | — | |
| CAGradBackbone=ConvNeXtV2-Base2025.09 | — | — | — | — | 0.3479 | 0.2579 | — | — | — | — | — | — | |
| GCondBackbone=ConvNeXtV2-Base2025.09 | — | — | — | — | 0.2677 | 0.1941 | — | — | — | — | — | — | |
| GradDropBackbone=ConvNeXtV2-Base2025.09 | — | — | — | — | 0.3173 | 0.2425 | — | — | — | — | — | — | |
| GradNormBackbone=ConvNeXtV2-Base2025.09 | — | — | — | — | 0.2888 | 0.1858 | — | — | — | — | — | — | |
| PCGradBackbone=ConvNeXtV2-Base2025.09 | — | — | — | — | 0.3166 | 0.24 | — | — | — | — | — | — |