Conditional Image Generation on ImageNet 256x256
4.98FIDReDi (VAE)
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| ReDi (VAE)Backbone=SiT-XL, Compression=VAE, KL weight (beta)=0.001, Training Iterations=400K, Representation Guidance Scale=optimal2025.12 | 4.98 | 4.55 | 61 | 77 | — | — | — | — | — | |
| ReDi (VAE)Backbone=SiT-XL, Compression=VAE, KL weight (beta)=0.01, Training Iterations=400K, Representation Guidance Scale=optimal2025.12 | 5.01 | 4.48 | 61 | 77 | — | — | — | — | — | |
| COT-FMModel=SiT-B/2, NFE=1002026.03 | 5.11 | — | — | — | — | — | — | — | — | |
| COT-FMModel=SiT-B/2, NFE=502026.03 | 5.28 | — | — | — | — | — | — | — | — | |
| ReDi (PCA)Backbone=SiT-XL, Compression=PCA, Training Iterations=400K, Representation Guidance Scale=optimal2025.12 | 5.48 | 4.66 | 59 | 77 | — | — | — | — | — | |
| Rectified FlowModel=SiT-B/2, NFE=1002026.03 | 5.82 | — | — | — | — | — | — | — | — | |
| Rectified FlowModel=SiT-B/2, NFE=502026.03 | 5.86 | — | — | — | — | — | — | — | — | |
| QuESTBit (W/A)=4/42025.12 | 5.98 | — | — | — | 202.45 | — | — | — | — | |
| HQ-DMBit (W/A)=4/42025.12 | 6.58 | 9.6 | 0.8444 | — | 248.3 | — | — | — | — | |
| EfficientDMBit (W/A)=4/42025.12 | 6.63 | 9.8 | 0.8097 | — | 220.2 | — | — | — | — | |
| EfficientDMBit (W/A)=8/42025.12 | 6.75 | 8.48 | 0.849 | — | 252 | — | — | — | — | |
| COT-FMModel=SiT-B/2, NFE=102026.03 | 7.52 | — | — | — | — | — | — | — | — | |
| COT-FMModel=SiT-B/4, NFE=1002026.03 | 7.65 | — | — | — | — | — | — | — | — | |
| HQ-DMBit (W/A)=8/42025.12 | 7.68 | 9.93 | 0.8868 | — | 292.76 | — | — | — | — | |
| COT-FMModel=SiT-B/4, NFE=502026.03 | 7.81 | — | — | — | — | — | — | — | — | |
| Rectified FlowModel=SiT-B/2, NFE=102026.03 | 8.25 | — | — | — | — | — | — | — | — | |
| Rectified FlowModel=SiT-B/4, NFE=1002026.03 | 8.3 | — | — | — | — | — | — | — | — | |
| Rectified FlowModel=SiT-B/4, NFE=502026.03 | 8.39 | — | — | — | — | — | — | — | — | |
| PTQDBit (W/A)=4/82025.12 | 8.74 | 7.98 | 0.9169 | — | 344.72 | — | — | — | — | |
| Q-DiffusionBit (W/A)=4/82025.12 | 9.29 | 9.29 | 0.9106 | — | 336.8 | — | — | — | — | |
| HQ-DMBit (W/A)=4/82025.12 | 9.85 | 8.37 | 0.9263 | — | 350.22 | — | — | — | — | |
| COT-FMModel=SiT-B/4, NFE=102026.03 | 9.87 | — | — | — | — | — | — | — | — | |
| EfficientDMBit (W/A)=4/82025.12 | 9.96 | 7.8 | 0.9248 | — | 351.97 | — | — | — | — | |
| PTQDBit (W/A)=8/82025.12 | 10.05 | 9.01 | 0.93 | — | 359.78 | — | — | — | — | |
| Q-DiffusionBit (W/A)=8/82025.12 | 10.6 | 9.29 | 0.9246 | — | 350.93 | — | — | — | — | |
| HQ-DMBit (W/A)=8/82025.12 | 11.01 | 7.76 | 0.9346 | — | 363.22 | — | — | — | — | |
| EfficientDMBit (W/A)=8/82025.12 | 11.02 | 7.65 | 0.938 | — | 363.71 | — | — | — | — | |
| Rectified FlowModel=SiT-B/4, NFE=102026.03 | 11.16 | — | — | — | — | — | — | — | — | |
| FPBit (W/A)=32/322025.12 | 11.22 | 7.78 | 0.9368 | — | 365.35 | — | — | — | — | |
| ReDi (VAE)Backbone=SiT-B, Compression=VAE, KL weight (beta)=0.01, Training Iterations=400K, Representation Guidance Scale=optimal2025.12 | 11.76 | 5.53 | 58 | 73 | — | — | — | — | — | |
| ReDi (VAE)Backbone=SiT-B, Compression=VAE, KL weight (beta)=0.001, Training Iterations=400K, Representation Guidance Scale=optimal2025.12 | 12.63 | 5.37 | 60 | 71 | — | — | — | — | — | |
| ReDi (PCA)Backbone=SiT-B, Compression=PCA, Training Iterations=400K, Representation Guidance Scale=optimal2025.12 | 18.49 | 6.33 | 58 | 65 | — | — | — | — | — | |
| HQ-DMBit (W/A)=4/32025.12 | 33.47 | 12.09 | 0.5487 | — | 49.62 | — | — | — | — | |
| NoiseRaterStep=+80k, Model Backbone=DiT-S/22026.05 | 48.65 | — | — | — | — | — | — | — | — | |
| NoiseRaterStep=+60k, Model Backbone=DiT-S/22026.05 | 49.11 | — | — | — | — | — | — | — | — | |
| Naive (max)Step=+80k, Model Backbone=DiT-S/22026.05 | 49.29 | — | — | — | — | — | — | — | — | |
| VanillaStep=+80k, Model Backbone=DiT-S/22026.05 | 49.59 | — | — | — | — | — | — | — | — | |
| NoiseRaterStep=+40k, Model Backbone=DiT-S/22026.05 | 49.88 | — | — | — | — | — | — | — | — | |
| Naive (max)Step=+60k, Model Backbone=DiT-S/22026.05 | 49.95 | — | — | — | — | — | — | — | — | |
| VanillaStep=+60k, Model Backbone=DiT-S/22026.05 | 50.11 | — | — | — | — | — | — | — | — | |
| Naive (max)Step=+40k, Model Backbone=DiT-S/22026.05 | 50.47 | — | — | — | — | — | — | — | — | |
| NoiseRaterStep=+20k, Model Backbone=DiT-S/22026.05 | 50.71 | — | — | — | — | — | — | — | — | |
| VanillaStep=+40k, Model Backbone=DiT-S/22026.05 | 50.88 | — | — | — | — | — | — | — | — | |
| Naive (max)Step=+20k, Model Backbone=DiT-S/22026.05 | 51.29 | — | — | — | — | — | — | — | — | |
| VanillaStep=+20k, Model Backbone=DiT-S/22026.05 | 51.42 | — | — | — | — | — | — | — | — | |
| Naive (min)Step=+80k, Model Backbone=DiT-S/22026.05 | 51.82 | — | — | — | — | — | — | — | — | |
| Naive (min)Step=+60k, Model Backbone=DiT-S/22026.05 | 52.52 | — | — | — | — | — | — | — | — | |
| Naive (min)Step=+40k, Model Backbone=DiT-S/22026.05 | 53.02 | — | — | — | — | — | — | — | — | |
| Naive (min)Step=+20k, Model Backbone=DiT-S/22026.05 | 53.28 | — | — | — | — | — | — | — | — | |
| EfficientDMBit (W/A)=4/32025.12 | 99.78 | 65.79 | 0.2748 | — | 8.74 | — | — | — | — | |
| COT-FMModel=SiT-B/2, NFE=22026.03 | 101.66 | — | — | — | — | — | — | — | — | |
| COT-FMModel=SiT-B/4, NFE=22026.03 | 114.1 | — | — | — | — | — | — | — | — | |
| Rectified FlowModel=SiT-B/2, NFE=22026.03 | 119.57 | — | — | — | — | — | — | — | — | |
| Rectified FlowModel=SiT-B/4, NFE=22026.03 | 134.99 | — | — | — | — | — | — | — | — | |
| COT-FMModel=SiT-B/2, NFE=12026.03 | 231.99 | — | — | — | — | — | — | — | — | |
| COT-FMModel=SiT-B/4, NFE=12026.03 | 241.18 | — | — | — | — | — | — | — | — | |
| Rectified FlowModel=SiT-B/2, NFE=12026.03 | 264.36 | — | — | — | — | — | — | — | — | |
| Rectified FlowModel=SiT-B/4, NFE=12026.03 | 276.13 | — | — | — | — | — | — | — | — | |
| FlexTok d18-d18Tokenizer Type=Diffusion, Tokenizer # Params=573M, Generator Model=Custom AR, Generator Type=AR, Generator # Params=1.33B, # Tokens=322026.06 | — | — | — | — | — | 1.61 | 2.02 | 7.45 | 0.46 | |
| FlexTok d18-d28Tokenizer Type=Diffusion, Tokenizer # Params=1.4B, Generator Model=Custom AR, Generator Type=AR, Generator # Params=1.33B, # Tokens=322026.06 | — | — | — | — | — | 1.45 | 1.86 | 27.52 | — | |
| Learnable Global Merging (100 steps)Tokenizer Type=ViT, Tokenizer # Params=176M, Generator Model=LightningDiT-XL, Generator Type=Diffusion, Generator # Params=675M, # Tokens=32, Sampling Steps=1002026.06 | — | — | — | — | 302.3 | 1.37 | 2.54 | 4.33 | 2.04 | |
| Learnable Global Merging (100 steps)Tokenizer Type=ViT, Tokenizer # Params=176M, Generator Model=LightningDiT-XL, Generator Type=Diffusion, Generator # Params=675M, # Tokens=128, Sampling Steps=1002026.06 | — | — | — | — | 290.7 | 0.59 | 1.86 | 11.51 | 0.36 | |
| Learnable Global Merging (25 steps)Tokenizer Type=ViT, Tokenizer # Params=176M, Generator Model=LightningDiT-XL, Generator Type=Diffusion, Generator # Params=675M, # Tokens=32, Sampling Steps=252026.06 | — | — | — | — | 311.2 | 1.37 | 2.89 | 1.08 | 6.12 | |
| Learnable Global Merging (25 steps)Tokenizer Type=ViT, Tokenizer # Params=176M, Generator Model=LightningDiT-XL, Generator Type=Diffusion, Generator # Params=675M, # Tokens=128, Sampling Steps=252026.06 | — | — | — | — | 297.4 | 0.59 | 2.03 | 2.88 | 1.95 | |
| MAETok-B-128Tokenizer Type=ViT, Tokenizer # Params=176M, Generator Model=LightningDiT-XL, Generator Type=Diffusion, Generator # Params=675M, # Tokens=1282026.06 | — | — | — | — | 308.4 | 0.48 | 1.73 | — | — | |
| One-D-Piece-B-256*Tokenizer Type=ViT, Tokenizer # Params=176M, Generator Model=MaskGIT-ViT, Generator Type=Masked, Generator # Params=177M, # Tokens=2562026.06 | — | — | — | — | 259.3 | 1.11 | 2.7 | — | — | |
| SD-VAE + DiT-XL/2Tokenizer Type=CNN, Tokenizer # Params=84M, Generator Model=DiT-XL/2, Generator Type=Diffusion, Generator # Params=675M, # Tokens=1024 (256)2026.06 | — | — | — | — | 278.2 | 0.62 | 2.27 | — | — | |
| SD-VAE + SiT-XL/2Tokenizer Type=CNN, Tokenizer # Params=84M, Generator Model=SiT-XL/2, Generator Type=Diffusion, Generator # Params=675M, # Tokens=1024 (256)2026.06 | — | — | — | — | 270.3 | 0.62 | 2.06 | — | — | |
| SD-VAE + SiT-XL/2 + REPATokenizer Type=CNN, Tokenizer # Params=84M, Generator Model=SiT-XL/2 + REPA, Generator Type=Diffusion, Generator # Params=675M, # Tokens=1024 (256)2026.06 | — | — | — | — | 305.7 | 0.62 | 1.42 | — | — | |
| Semanticist-LTokenizer Type=Diffusion, Tokenizer # Params=609M, Generator Model=ϵLlamaGen-L, Generator Type=MAR, Generator # Params=489M, # Tokens=322026.06 | — | — | — | — | 260.9 | 1.68 | 2.57 | 11.18 | 0.015 | |
| Semanticist-XLTokenizer Type=Diffusion, Tokenizer # Params=827M, Generator Model=ϵLlamaGen-L, Generator Type=MAR, Generator # Params=489M, # Tokens=322026.06 | — | — | — | — | 254 | 1.4 | 2.57 | 16.09 | — | |
| SoftVQ-B-32Tokenizer Type=ViT, Tokenizer # Params=176M, Generator Model=SiT-XL, Generator Type=Diffusion, Generator # Params=675M, # Tokens=322026.06 | — | — | — | — | 301.3 | 0.89 | 2.51 | 10.82 | 0.61 | |
| TiTok-B-64Tokenizer Type=ViT, Tokenizer # Params=176M, Generator Model=MaskGIT-ViT, Generator Type=Masked, Generator # Params=177M, # Tokens=642026.06 | — | — | — | — | 216.6 | 1.7 | 2.48 | — | — | |
| VA-VAETokenizer Type=CNN, Tokenizer # Params=70M, Generator Model=LightningDiT-XL, Generator Type=Diffusion, Generator # Params=675M, # Tokens=2562026.06 | — | — | — | — | 295.3 | 0.28 | 1.35 | — | — |