Class-conditional Image Generation on ImageNet-1k (train)
1.13FIDGAE
Evaluation Results
| Method | Links | |
|---|---|---|
| GAE# Params=675M, NFE=250 × 2, setup=Multi-NFE diffusion/flow2026.05 | 1.13 | |
| RAE + DiT-XL# Params=839M, NFE=50 × 2, setup=Multi-NFE diffusion/flow2026.05 | 1.13 | |
| REPA-E# Params=675M, NFE=250 × 2, setup=Multi-NFE diffusion/flow2026.05 | 1.15 | |
| LightningDiT-XL/2# Params=675M, NFE=250 × 2, setup=Multi-NFE diffusion/flow2026.05 | 1.35 | |
| SiT-XL/2 + REG# Params=675M, NFE=250 × 2, setup=Multi-NFE diffusion/flow2026.05 | 1.36 | |
| SiT-XL/2 + REPA# Params=675M, NFE=250 × 2, setup=Multi-NFE diffusion/flow2026.05 | 1.42 | |
| mmDiT-XL/2 (ours)Resolution=2562024.11 | 1.59 | |
| DiT-XL/2 w/AttImplementation=Our re-implementation, Resolution=2562024.11 | 1.71 | |
| mmDiT-XL/2 (SD3)Implementation=Our re-implementation, Resolution=2562024.11 | 1.71 | |
| iMF-XL/2# Params=610M, NFE=1, setup=1-NFE diffusion/flow from scratch2026.05 | 1.72 | |
| mDT-v2-XL/2 w/ LNResolution=2562024.11 | 1.79 | |
| DiT-XL/2 w/LNImplementation=Our re-implementation, Resolution=2562024.11 | 1.95 | |
| UNet (SDXL)Implementation=Our re-implementation, Resolution=2562024.11 | 2.05 | |
| PixArt-a-XL/2Implementation=Our re-implementation, Resolution=2562024.11 | 2.06 | |
| SiT-XL/2# Params=675M, NFE=250 × 2, setup=Multi-NFE diffusion/flow2026.05 | 2.06 | |
| Sphere Latent Encoder (Ours-XL/2)# Params=675M, NFE=6 × 2, setup=Sphere model2026.05 | 2.11 | |
| Sphere Latent Encoder (Ours-XL/2)# Params=675M, NFE=4 × 2, setup=Sphere model2026.05 | 2.25 | |
| DiT-XL2 w/ LNResolution=2562024.11 | 2.27 | |
| DiT-XL/2# Params=675M, NFE=250 × 2, setup=Multi-NFE diffusion/flow2026.05 | 2.27 | |
| mDT-v2-XL/2 w/ LNImplementation=Our re-implementation, Resolution=2562024.11 | 2.51 | |
| α-Flow-XL/2+# Params=676M, NFE=1, setup=1-NFE diffusion/flow from scratch2026.05 | 2.58 | |
| mmDiT-XL/2 (ours)Resolution=5122024.11 | 2.76 | |
| DiT-XL/2 w/LNImplementation=Our re-implementation, Resolution=5122024.11 | 2.85 | |
| mmDiT-XL/2 (SD3)Implementation=Our re-implementation, Resolution=5122024.11 | 3.02 | |
| DiT-XL2 w/ LNResolution=5122024.11 | 3.04 | |
| PixArt-a-XL/2Implementation=Our re-implementation, Resolution=5122024.11 | 3.05 | |
| MeanFlow-XL/2# Params=676M, NFE=1, setup=1-NFE diffusion/flow from scratch2026.05 | 3.43 | |
| UNet (SD/LDM-G4)Resolution=2562024.11 | 3.6 | |
| mDT-v2-XL/2 w/ LNImplementation=Our re-implementation, Resolution=5122024.11 | 3.75 | |
| Sphere Encoder# Params=1.3B, NFE=4 × 2, setup=Sphere model2026.05 | 4.02 | |
| UNet (SDXL)Implementation=Our re-implementation, Resolution=5122024.11 | 4.81 |