Image Generation on LSUN Bedroom 256x256 (test)
1.88FIDTDPM+ (TTrunc=99)
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| TDPM+ (TTrunc=99)Backbone=ADM, NFE=100, TTrunc=992022.02 | 1.88 | — | — | — | — | |
| TDPM+ (TTrunc=49)Backbone=ADM, NFE=50, TTrunc=492022.02 | 1.89 | — | — | — | — | |
| ADM#Param=552.8M, MACs=1114.2G2022.11 | 1.9 | — | — | — | — | |
| ADMBackbone=ADM, NFE=10002022.02 | 1.9 | — | — | — | — | |
| TDPM (TTrunc=99)Backbone=ADM, NFE=100, TTrunc=992022.02 | 2.24 | — | — | — | — | |
| ADM+DDIMBackbone=ADM, NFE=2502022.02 | 2.31 | — | — | — | — | |
| StyleGAN2020.06 | 2.65 | — | — | — | — | |
| TDPM (TTrunc=49)Backbone=ADM, NFE=50, TTrunc=492022.02 | 2.92 | — | — | — | — | |
| LDM#Param=274.1M, MACs=96.1G2022.11 | 3 | — | — | — | — | |
| Full precisionBitwidth (W/A)=32/32, model=LDM-4, steps=200, eta=1.02023.05 | 3 | — | 7.13 | — | — | |
| FPPrec. (W/A)=32/32, Architecture=LDM-4, Steps=100, eta=12025.05 | 3.02 | — | — | 2.29 | — | |
| Diffusion StyleGAN22022.06 | 3.65 | 32 | — | — | — | |
| TDPM+ (TTrunc=99)Backbone=DDPM, NFE=100, TTrunc=992022.02 | 3.67 | — | — | — | — | |
| PTQDBitwidth (W/A)=8/8, model=LDM-4, steps=200, eta=1.02023.05 | 3.75 | — | 9.89 | — | — | |
| Q-DiffusionBitwidth (W/A)=8/8, model=LDM-4, steps=200, eta=1.02023.05 | 3.8 | — | 9.95 | — | — | |
| TDPM (TTrunc=99)Backbone=DDPM, NFE=100, TTrunc=992022.02 | 3.95 | — | — | — | — | |
| TDPM+ (TTrunc=49)Backbone=DDPM, NFE=50, TTrunc=492022.02 | 3.98 | — | — | — | — | |
| StyleGAN2Backbone=GAN-based, NFE=12022.02 | 3.98 | — | — | — | — | |
| StyleGAN22022.06 | 3.98 | 32 | — | — | — | |
| TDPM+ (TTrunc=4)Backbone=ADM, NFE=5, TTrunc=42022.02 | 4.01 | — | — | — | — | |
| TDPM+ (TTrunc=4)Backbone=DDPM, NFE=5, TTrunc=42022.02 | 4.09 | — | — | — | — | |
| TDPM (TTrunc=49)Backbone=DDPM, NFE=50, TTrunc=492022.02 | 4.1 | — | — | — | — | |
| TDPM (TTrunc=4)Backbone=DDPM, NFE=5, TTrunc=42022.02 | 4.16 | — | — | — | — | |
| IDDPM#Param=113.7M, MACs=248.6G2022.11 | 4.2 | — | — | — | — | |
| StyleGAN2 + DiffAug2022.06 | 4.25 | 19 | — | — | — | |
| Denoising Diffusion Probabilistic Modelsloss_objective=L_simple, model_scale=large2020.06 | 4.9 | — | — | — | — | |
| DDPM#Param=113.7M, MACs=248.7G2022.11 | 4.9 | — | — | — | — | |
| DDPMBackbone=DDPM, NFE=10002022.02 | 4.9 | — | — | — | — | |
| SD#Param=21.1M, MACs=6.7G2022.11 | 5.2 | — | — | — | — | |
| TDPM (TTrunc=0)Backbone=GAN-based, NFE=1, TTrunc=02022.02 | 5.24 | — | — | — | — | |
| PTQDBitwidth (W/A)=MP, model=LDM-4, steps=200, eta=1.02023.05 | 5.49 | — | 12.04 | — | — | |
| Q-DiffusionBitwidth (W/A)=MP, model=LDM-4, steps=200, eta=1.02023.05 | 5.75 | — | 12.79 | — | — | |
| PTQDBitwidth (W/A)=4/8, model=LDM-4, steps=200, eta=1.02023.05 | 5.94 | — | 15.16 | — | — | |
| COCO-GANModel Parameters=N2,M2,S32, Convergence=Not fully converged2019.03 | 5.99 | — | — | — | — | |
| FPModel=LDM-4, Steps=50, W/A=32/322025.05 | 6.17 | 51.46 | 13.92 | — | 64.3 | |
| Denoising Diffusion Probabilistic Modelsloss_objective=L_simple2020.06 | 6.36 | — | — | — | — | |
| DDIMBackbone=DDPM, NFE=502022.02 | 6.62 | — | — | — | — | |
| Q-DiffusionBitwidth (W/A)=4/8, model=LDM-4, steps=200, eta=1.02023.05 | 6.72 | — | 18.8 | — | — | |
| FPBit (W/A)=32/32, Model architecture=LDM-4, Denoising steps=1002025.12 | 7.56 | 41.18 | 14.06 | — | 51.95 | |
| StyleGAN2-ADABackbone=GAN-based, NFE=12022.02 | 7.89 | — | — | — | — | |
| StyleGAN2 + ADA2022.06 | 7.89 | 5 | — | — | — | |
| TDPM (TTrunc=4)Backbone=ADM, NFE=5, TTrunc=42022.02 | 7.92 | — | — | — | — | |
| PGGAN2019.03 | 8.34 | — | — | — | — | |
| ProgressiveGAN2020.06 | 8.34 | — | — | — | — | |
| TALORAPrec. (W/A)=6/6, h=4, Architecture=LDM-4, Steps=100, eta=12025.05 | 8.4 | — | — | 2.49 | — | |
| TALORAPrec. (W/A)=6/6, h=2, Architecture=LDM-4, Steps=100, eta=12025.05 | 8.42 | — | — | 2.49 | — | |
| SA-QLoRAModel=LDM-4, Steps=50, W/A=4/82025.05 | 8.79 | 52.31 | 12.41 | — | 50.16 | |
| IntroVAE2019.03 | 8.84 | — | — | — | — | |
| CacheQuantCache Frequency (N)=2, Precision=W8A8, Bops=3.26T, Speedup=3.05x, Compression Ratio=3.99x, Retrain=false, T=100, eta=02025.03 | 8.85 | — | — | — | — | |
| CacheQuantCache Frequency (N)=3, Precision=W8A8, Bops=2.34T, Speedup=4.72x, Compression Ratio=3.99x, Retrain=false, T=100, eta=02025.03 | 9.27 | — | — | — | — | |
| EfficientDMModel=LDM-4, Steps=50, W/A=4/82025.05 | 9.32 | 53.02 | 14.48 | — | 48.62 | |
| SA-PTQModel=LDM-4, Steps=50, W/A=8/82025.05 | 9.83 | 54.54 | 13.85 | — | 56.86 | |
| Q-DiffusionPrec. (W/A)=6/6, Architecture=LDM-4, Steps=100, eta=12025.05 | 10.1 | — | — | 2.11 | — | |
| QuESTPrec. (W/A)=6/6, Architecture=LDM-4, Steps=100, eta=12025.05 | 10.1 | — | — | 2.2 | — | |
| CacheQuantCache Frequency (N)=5, Precision=W8A8, Bops=1.54T, Speedup=7.06x, Compression Ratio=3.99x, Retrain=false, T=100, eta=02025.03 | 10.29 | — | — | — | — | |
| PTQDModel=LDM-4, Steps=50, W/A=8/82025.05 | 10.31 | 54.15 | 15.89 | — | 56.57 | |
| LDM-4Bops=98.36T, Speed=1.00x, Size=1317.4, Retrain=false, T=100, eta=02025.03 | 10.49 | — | — | — | — | |
| EDA-DMPrec. (W/A)=6/6, Architecture=LDM-4, Steps=100, eta=12025.05 | 10.56 | — | — | 2.12 | — | |
| Lite-LDM#Param=22.4M, MACs=7.9G2022.11 | 10.9 | — | — | — | — | |
| PTQDModel=LDM-4, Steps=50, W/A=4/82025.05 | 10.96 | 52.8 | 15.42 | — | 47.83 | |
| DeepcacheCache Frequency (N)=2, Bops=52.23T, Speed=1.79x, Compression Ratio=1.00x, Retrain=false, T=100, eta=02025.03 | 11.21 | — | — | — | — | |
| DeepcacheCache Frequency (N)=3, Bops=37.49T, Speed=2.68x, Compression Ratio=1.00x, Retrain=false, T=100, eta=02025.03 | 11.86 | — | — | — | — | |
| TALORAPrec. (W/A)=4/4, h=2, Architecture=LDM-4, Steps=100, eta=12025.05 | 12.21 | — | — | 2.47 | — | |
| TALORAPrec. (W/A)=4/4, h=4, Architecture=LDM-4, Steps=100, eta=12025.05 | 12.34 | — | — | 2.48 | — | |
| HQ-DMBit (W/A)=8/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 12.35 | 37.44 | 16.27 | — | 47.31 | |
| EfficientDMPrec. (W/A)=6/6, Architecture=LDM-4, Steps=100, eta=12025.05 | 12.95 | — | — | 2.57 | — | |
| HQ-DMBit (W/A)=4/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 13.15 | 36.34 | 17.53 | — | 45.32 | |
| DeepcacheCache Frequency (N)=5, Bops=24.59T, Speed=4.08x, Compression Ratio=1.00x, Retrain=false, T=100, eta=02025.03 | 14.28 | — | — | — | — | |
| SA-QLoRAModel=LDM-4, Steps=50, W/A=4/42025.05 | 14.72 | 45.6 | 18.19 | — | 41.84 | |
| EfficientDMBit (W/A)=4/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 16.81 | 29.66 | 20.81 | — | 41.42 | |
| EfficientDMBit (W/A)=8/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 17.91 | 37.86 | 19.44 | — | 40.4 | |
| HQ-DMBit (W/A)=3/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 19.22 | 31.92 | 19.29 | — | 38.32 | |
| EfficientDMModel=LDM-4, Steps=50, W/A=4/42025.05 | 19.3 | 40.09 | 22.63 | — | 43.77 | |
| EfficientDMBit (W/A)=3/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 20.92 | 29.38 | 22.78 | — | 38.57 | |
| HQ-DMBit (W/A)=4/3, Model architecture=LDM-4, Denoising steps=1002025.12 | 21.88 | 29.1 | 22.15 | — | 34.52 | |
| TFMQ-DMBit (W/A)=3/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 25.74 | — | 35.18 | — | 32.2 | |
| TFMQ-DMBit (W/A)=2/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 25.77 | — | 36.74 | — | — | |
| EfficientDMBit (W/A)=4/3, Model architecture=LDM-4, Denoising steps=1002025.12 | 26.8 | 29.02 | 24.01 | — | 33.43 | |
| HQ-DMBit (W/A)=2/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 28.5 | 18.76 | 30.17 | — | 32.64 | |
| EfficientDMBit (W/A)=2/4, Model architecture=LDM-4, Denoising steps=1002025.12 | 33.68 | 20.34 | 27.19 | — | 28.76 | |
| EfficientDMPrec. (W/A)=4/4, Architecture=LDM-4, Steps=100, eta=12025.05 | 36.36 | — | — | 2.69 | — |