Image Generation on ImageNet-1K 256x256
0.99FIDBAR-L
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| BAR-L#params=1.1B, precision=float32, hardware=single H200, KV-cache=true2026.02 | 0.99 | 10.65 | — | — | — | — | |
| RAE#params=839M, precision=float32, hardware=single H2002026.02 | 1.13 | 6.62 | — | — | — | — | |
| BAR-B#params=415M, precision=float32, hardware=single H200, KV-cache=true2026.02 | 1.13 | 24.33 | — | — | — | — | |
| xAR#params=1.1B, precision=float32, hardware=single H2002026.02 | 1.24 | 2.03 | — | — | — | — | |
| DDT#params=675M, precision=float32, hardware=single H2002026.02 | 1.26 | 1.62 | — | — | — | — | |
| BAR-B/2#params=415M, precision=float32, hardware=single H200, KV-cache=true2026.02 | 1.35 | 150.52 | — | — | — | — | |
| VA-VAE#params=675M, precision=float32, hardware=single H2002026.02 | 1.35 | 1.51 | — | — | — | — | |
| RAR-XLModel Type=Raster-scan AR, # Params=955M2025.11 | 1.5 | — | 306.9 | — | 80 | 62 | |
| MAR#params=943M, precision=float32, hardware=single H2002026.02 | 1.55 | 1.19 | — | — | — | — | |
| STAR-XLModel Type=Randomized AR, # Params=955M2025.11 | 1.65 | — | 333.2 | — | 80 | 62 | |
| RAR-LModel Type=Raster-scan AR, # Params=461M2025.11 | 1.7 | — | 299.5 | — | 81 | 60 | |
| VAR#params=2.0B, precision=float32, hardware=single H2002026.02 | 1.92 | 8.08 | — | — | — | — | |
| STAR-LModel Type=Randomized AR, # Params=461M2025.11 | 1.98 | — | 322 | — | 82 | 58 | |
| MeanFlow#params=676M, precision=float32, hardware=single H2002026.02 | 2.2 | 151.48 | — | — | — | — | |
| RandAR-XLModel Type=Randomized AR, # Params=775M2025.11 | 2.25 | — | 314.2 | — | 80 | 60 | |
| DiT-XL/2Model Type=Diffusion, # Params=675M2025.11 | 2.27 | — | 278.2 | — | 83 | 57 | |
| PAR-4x#params=3.1B, precision=float32, hardware=single H2002026.02 | 2.29 | 4.92 | — | — | — | — | |
| BAR-B/4#params=416M, precision=float32, hardware=single H200, KV-cache=true2026.02 | 2.34 | 445.48 | — | — | — | — | |
| RandAR-LModel Type=Randomized AR, # Params=343M2025.11 | 2.55 | — | 288.8 | — | 81 | 58 | |
| GADCData Budget=0.8% (10K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 3.43 | — | — | — | — | — | |
| D2CData Budget=0.8% (10K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 4.2 | — | — | — | — | — | |
| FSF-DMDBackbone=LightningDiT-B/1, Training steps=400K, Training protocol=from scratch2026.05 | 6.35 | — | 194.91 | — | — | — | |
| iMF BaselineBackbone=LightningDiT-B/1, Training steps=400K, Training protocol=from scratch2026.05 | 9.1 | — | 148.71 | — | — | — | |
| GADCData Budget=4.0% (50K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 11.01 | — | — | — | — | — | |
| RCGMBackbone=LightningDiT-B/1, Training steps=400K, Training protocol=from scratch2026.05 | 11.53 | — | 164.55 | — | — | — | |
| iMF + TwinFlowBackbone=LightningDiT-B/1, Training steps=400K, Training protocol=from scratch2026.05 | 12.17 | — | 144.26 | — | — | — | |
| iMF + DMDBackbone=LightningDiT-B/1, Training steps=400K, Training protocol=from scratch, TTUR=1:52026.05 | 13.38 | — | 138.8 | — | — | — | |
| D2CData Budget=4.0% (50K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 14.81 | — | — | — | — | — | |
| RobustQuantizerW/A=4/4, Granularity=Channel, Calib.=✗, Steps=50, CFG=4.02025.09 | 16.13 | — | — | — | 92.91 | — | |
| SVDQuantW/A=4/4, Granularity=Group=64, Calib.=✓, Steps=50, CFG=4.02025.09 | 16.35 | — | — | — | 90.87 | — | |
| GADCData Budget=8.0% (100K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 17.09 | — | — | — | — | — | |
| FP (32/32)W/A=32/32, Granularity=-, Calib.=-, Steps=50, CFG=4.02025.09 | 19.11 | — | — | — | 92.98 | — | |
| D2CData Budget=8.0% (100K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 22.55 | — | — | — | — | — | |
| RCGM + TwinFlowBackbone=LightningDiT-B/1, Training steps=400K, Training protocol=from scratch2026.05 | 22.58 | — | 106.04 | — | — | — | |
| HerdingData Budget=4.0% (50K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 32.38 | — | — | — | — | — | |
| SiT-B/2 + Data WarmupModel backbone=SiT-B/2, Data sampling/warmup protocol=Data Warmup, Resolution=256x2562026.04 | 32.75 | — | 45.7 | 6.56 | 54 | 63 | |
| RandomData Budget=0.8% (10K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 35.86 | — | — | — | — | — | |
| SiT-B/2Model backbone=SiT-B/2, Data sampling/warmup protocol=uniform data sampling, Resolution=256x2562026.04 | 36.16 | — | 41.4 | 6.8 | 52 | 63 | |
| HerdingData Budget=8.0% (100K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 36.37 | — | — | — | — | — | |
| RandomData Budget=4.0% (50K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 36.78 | — | — | — | — | — | |
| HerdingData Budget=0.8% (10K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 40.75 | — | — | — | — | — | |
| RandomData Budget=8.0% (100K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 41.02 | — | — | — | — | — | |
| SiT-B/2 + Inverse Data WarmupModel backbone=SiT-B/2, Data sampling/warmup protocol=Inverse Data Warmup, Resolution=256x2562026.04 | 41.05 | — | 36.6 | 7.19 | 49 | 62 | |
| K-CenterData Budget=0.8% (10K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 50.77 | — | — | — | — | — | |
| K-CenterData Budget=4.0% (50K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 69.86 | — | — | — | — | — | |
| K-CenterData Budget=8.0% (100K), Backbone=DiT-L/2, Training Iterations=100k, Resolution=256x2562026.06 | 71.31 | — | — | — | — | — | |
| QueSTW/A=4/4, Granularity=Channel, Calib.=✓, Steps=50, CFG=4.02025.09 | 84.2 | — | — | — | 26.51 | — | |
| SVDQuantW/A=4/4, Granularity=Channel, Calib.=✓, Steps=50, CFG=4.02025.09 | 282.19 | — | — | — | 0.98 | — |