Image Generation on 1024x1024
145.6Latency (ms)BADiff
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| BADiffbpp=0.8–1.22025.10 | 145.6 | — | — | — | — | — | — | — | 17.8 | — | |
| PNDM+LICbpp=0.8–1.22025.10 | 192.5 | — | — | — | — | — | — | — | 20.1 | — | |
| DDPM+LICbpp=0.8–1.22025.10 | 228.7 | — | — | — | — | — | — | — | 21.5 | — | |
| SparVAR#Scales=14, Resolution=1024x10242026.02 | 383.11 | 1.16 | — | 19.29 | 26.98 | 21.157 | 0.735 | 0.24 | — | — | |
| FastVAR#Scales=14, Resolution=1024x10242026.02 | 403.33 | 1.11 | — | 19.33 | 25.72 | 20.85 | 0.704 | 0.316 | — | — | |
| HART#Scales=14, FlashAttention=true, Resolution=1024x10242026.02 | 446.3 | 1 | — | 19.35 | 27.93 | — | — | — | — | — | |
| Fused INT8Variant=Fused INT8, Steps=48, Resolution=1024px, Hardware=RTX 3090, Note=peak 23.40 GB, 4/4 images2026.06 | 156,490 | — | — | — | — | — | — | — | — | 1 | |
| NF4Variant=NF4, Steps=48, Resolution=1024px, Hardware=RTX 30902026.06 | 164,500 | — | — | — | — | — | — | — | — | 1 | |
| FP8Variant=FP8, Steps=48, Resolution=1024px, Hardware=RTX 3090, Note=dequant-to-bf16 on Ampere2026.06 | 172,900 | — | — | — | — | — | — | — | — | 2 | |
| INT8 fake-quantVariant=INT8 fake-quant, Steps=48, Resolution=1024px, Hardware=RTX 3090, Note=quantize then dequant to bf162026.06 | 184,000 | — | — | — | — | — | — | — | — | 2 |