Text-to-Image Generation on DPG-Bench (DPG Metric)
88.9DPG ScoreBase model (Qwen-Image)
Evaluation results
| Method | Links | |
|---|---|---|
| Base model (Qwen-Image)Inference Latency (ms)=24312026.02 | 88.9 | |
| ArcFlow-QwenNFE=22026.02 | 88.46 | |
| Qwen-Image-LightningNFE=22026.02 | 88.42 | |
| Qwen-Image-20BNFE=50 x 22026.02 | 88.32 | |
| Qwen-ImageInteraction Type=Self-Attn, Model Parameters=20B2025.11 | 88.32 | |
| Qwen-ImageType=Uni.2025.10 | 88.32 | |
| Z-ImageType=Uni.2025.10 | 88.14 | |
| E-DiT-baseInference Latency (ms)=17022026.02 | 88.1 | |
| PPCLInference Latency (ms)=17922026.02 | 87.9 | |
| Qwen-ImageDiffusion Model Type=Latent, Resolution=1024x10242026.05 | 87.8 | |
| TwinFlow (Qwen)NFE=22026.02 | 87.01 | |
| MoS-LInteraction Type=MoS, Model Parameters=5B2025.11 | 87.01 | |
| PG 3Parameters=24B2025.05 | 87 | |
| Ming-Flash-OmniType=Uni.2025.10 | 86.98 | |
| MeiGen-Image-FaQBackbone=MeiGen-Image, Framework=Forge-and-Quench2026.01 | 86.83 | |
| AsymFLUX.2 kleinDiffusion Model Type=Pixel, Resolution=1024x10242026.05 | 86.8 | |
| pi-Flow (GM-Qwen)NFE=22026.02 | 86.45 | |
| MoS-SInteraction Type=MoS, Model Parameters=3B2025.11 | 86.33 | |
| MeiGen-ImageBackbone=MeiGen-Image, Framework=Baseline2026.01 | 85.94 | |
| E-DiT-turboInference Latency (ms)=12832026.02 | 85.4 | |
| FLUX.2 klein BaseDiffusion Model Type=Latent, Resolution=1024x10242026.05 | 85.2 | |
| Show-o2Size=2B2026.02 | 85.02 | |
| Sana 1.0 1.6BParameters=1.6B2025.05 | 84.8 | |
| Sana#Params=1.6B2026.02 | 84.8 | |
| UniFlowModel Size=0.6B, Type=Diffusion, Res.=2562025.10 | 84.76 | |
| SANA-1.5Interaction Type=Cross-Attn, Model Parameters=4.8B2025.11 | 84.7 | |
| Tar-7B W/Self-ReflectParams=7B, Type=AR, #Training Images=46M, Res.=2562025.03 | 84.65 | |
| NExT-OmniSize=7B2026.02 | 84.46 | |
| SANAModel Size=0.6B, Type=Diffusion, Res.=5122025.10 | 84.3 | |
| ArcFlow-FLUXNFE=22026.02 | 84.29 | |
| PS-VAE32cStride=16, Channel Dimension=322025.12 | 84.25 | |
| Janus-Pro-7BModel Size=7B2026.01 | 84.2 | |
| Janus-Pro-7BStage=Uni. 1 Stage2025.12 | 84.19 | |
| Tar-7BParams=7B, Type=AR, #Training Images=46M, Res.=2562025.03 | 84.19 | |
| JanusPro-7BType=Uni.2025.10 | 84.19 | |
| FLUX.1-devNFE=502026.02 | 84.16 | |
| SD 3 MParameters=2B, Training Data=1B2025.05 | 84.1 | |
| Tar-1.5B W/Self-ReflectParams=3B, Type=AR, #Training Images=46M, Res.=2562025.03 | 84.1 | |
| SD3-MediumType=Gen.2025.10 | 84.08 | |
| FLUX.1 [dev]Parameters=12B2025.05 | 84 | |
| PPF-FLUX.1Resolution=1024, Testing FLOPs (%)=59.12025.06 | 84 | |
| FLUX.1 devDiffusion Model Type=Latent, Resolution=1024x10242026.05 | 84 | |
| FLUX.1-ftResolution=1024, Testing FLOPs (%)=1002025.06 | 83.89 | |
| FLUX.1[Dev]Interaction Type=Self-Attn, Model Parameters=12B2025.11 | 83.84 | |
| Ovis-U1Size=3B2026.02 | 83.72 | |
| FLUX.1-devBackbone=FLUX.1-dev, Framework=Baseline2026.01 | 83.66 | |
| PS-VAE96cStride=16, Channel Dimension=962025.12 | 83.62 | |
| Sana#Params=0.6B2026.02 | 83.6 | |
| SemHiTokParams=7B, Type=AR, #Training Images=15M, Res.=2562025.03 | 83.59 | |
| OmniGen2Size=3B2026.02 | 83.57 | |
| OmniGen2Type=Uni.2025.10 | 83.57 | |
| DALL-E 32025.05 | 83.5 | |
| DALL-E 3Type=Diff, Res.=10242025.03 | 83.5 | |
| PixelDiT-T2IDiffusion Model Type=Pixel, Resolution=1024x10242026.05 | 83.5 | |
| LiquidParams=7B, Type=AR, #Training Images=30M, Res.=5122025.03 | 83.45 | |
| UniTokParams=7B, Type=AR, #Training Images=30M, Res.=2562025.03 | 83.45 | |
| HPInference Latency (ms)=17862026.02 | 83.3 | |
| MAR-VAEStride=162025.12 | 83.19 | |
| FLUX.1Resolution=1024, Testing FLOPs (%)=1002025.06 | 83.14 | |
| PPF-FLUX.1Resolution=2048, Testing FLOPs (%)=53.92025.06 | 83.1 | |
| FLUX.1-ftResolution=2048, Testing FLOPs (%)=1002025.06 | 83.02 | |
| FLUX.1-dev-FaQBackbone=FLUX.1-dev, Framework=Forge-and-Quench2026.01 | 83.01 | |
| Tar-1.5BParams=3B, Type=AR, #Training Images=46M, Res.=2562025.03 | 82.96 | |
| PPF-FLUX.1Resolution=512, Testing FLOPs (%)=62.22025.06 | 82.9 | |
| FLUX.1-ftResolution=512, Testing FLOPs (%)=1002025.06 | 82.87 | |
| Tar-1.5BResolution=512, Model Size=1.5B2026.01 | 82.8 | |
| Tar-1.5B + ZipAR-16Resolution=512, Model Size=1.5B2026.01 | 82.8 | |
| Tar-1.5B + EAGLE-2Resolution=512, Model Size=1.5B2026.01 | 82.8 | |
| FLUX.1Resolution=2048, Testing FLOPs (%)=1002025.06 | 82.75 | |
| Janus-Pro-1BStage=Uni. 1 Stage2025.12 | 82.63 | |
| Janus-Pro-1BModel Size=1B2026.01 | 82.6 | |
| Harmon-1.5BModel Size=1.5B2026.01 | 82.6 | |
| FLUX.1Resolution=512, Testing FLOPs (%)=1002025.06 | 82.51 | |
| Tar-1.5B + ZipAR-16Resolution=1024, Model Size=1.5B2026.01 | 82.5 | |
| VAVAEStride=162025.12 | 82.45 | |
| Tar-1.5B + MuLo-SDResolution=512, Model Size=1.5B, Parallelism=2x2026.01 | 82.4 | |
| Sparse-LaViDaSFT=true, Resolution=10242025.12 | 82.4 | |
| Pi-Flow (GM-FLUX)NFE=22026.02 | 82.36 | |
| Tar-1.5BResolution=1024, Model Size=1.5B2026.01 | 82.3 | |
| Tar-1.5B + EAGLE-2Resolution=1024, Model Size=1.5B2026.01 | 82.3 | |
| Tar-1.5B + LANTERNResolution=1024, Model Size=1.5B2026.01 | 82.3 | |
| RAEv2Model=DiTDH-XL, Params=0.9B, Protocol=Finetuning2026.05 | 82.3 | |
| RAEv2Training Stage=Finetuning2026.05 | 82.3 | |
| Ming-UniVisionParams=16B-A3B, Type=AR-Continuous, Res.=5122025.03 | 82.12 | |
| Tar-1.5B + LANTERNResolution=512, Model Size=1.5B2026.01 | 82.1 | |
| LaViDa-O*SFT=true, Resolution=10242025.12 | 82.1 | |
| Tar-1.5B + MuLo-SDResolution=1024, Model Size=1.5B, Parallelism=4x2026.01 | 82 | |
| LaViDa-OSFT=false, Resolution=10242025.12 | 81.8 | |
| Ming-Lite-OmniSize=7B2026.02 | 81.72 | |
| RAEStride=162025.12 | 81.72 | |
| RAEv2Model=DiTDH-XL, Params=0.9B, Protocol=Pretraining2026.05 | 81.7 | |
| RAEv2Training Stage=Pretraining2026.05 | 81.7 | |
| FuseDiTParameters=2B, Training Data=26M2025.05 | 81.6 | |
| BLIP3o#Params=8B2026.02 | 81.6 | |
| UniWorld-V1Type=Uni.2025.10 | 81.38 | |
| CSFM#Params=1.3B, Backbone=UnifiedNextDiT2026.02 | 81.11 | |
| MetaQuery-LStage=Uni. 2 Stage2025.12 | 81.1 | |
| UnisonStage=Uni. 2 Stage2025.12 | 81.01 | |
| TinyFusionInference Latency (ms)=17892026.02 | 80.7 | |
| Emu3Model Size=≤2B2026.01 | 80.6 |