List your company

Class-conditional Image Generation on ImageNet 512x512 (val)

1.25FID (Val)EDM2-XXL + Autoguidance
-0.746412.729326.20539.6807Jan 26, 2023Aug 16, 2023Mar 5, 2024Sep 24, 2024Apr 14, 2025Nov 2, 2025May 24, 2026

Evaluation results

MethodLinks
EDM2-XXL + Autoguidanceguiding_model_capacity=M, training_amount=T/3.5, guidance_weight=2.05, EMA_length_main=0.075, EMA_length_guiding=0.1552024.061.25————————————
EDM2-S + Autoguidanceguiding_model_capacity=XS, training_amount=T/16, guidance_weight=2.1, EMA_length_main=0.07, EMA_length_guiding=0.1252024.061.34————————————
EDM2-XXL + Guidance intervalguidance_weight=2, EMA_length_main=0.015, EMA_length_guiding=0.0152024.061.4————————————
EDM2-S + Autoguidance (Reduce training only)guiding_model_capacity=XS, training_amount=T/16, guidance_weight=2.2, EMA_length_main=0.09, EMA_length_guiding=0.132024.061.51————————————
EDM2-S + Autoguidance (Same EMA for both)guiding_model_capacity=XS, training_amount=T/16, guidance_weight=1.95, EMA_length_main=0.05, EMA_length_guiding=0.052024.061.53————————————
MAR-L + l-DeTok#params=479M, Training Protocol=fine-tune, 200ep2025.071.61—315.7——————————
EDM2-S + Guidance intervalguidance_weight=2.1, EMA_length_main=0.025, EMA_length_guiding=0.0252024.061.68————————————
MAR-L + MAR-VAE#params=479M2025.071.73—279.9——————————
EDM2-XXL + Classifier-free guidanceguidance_weight=1.2, EMA_length_main=0.015, EMA_length_guiding=0.0152024.061.81————————————
BA-solverNumber of Function Evaluations (NFE)=152026.011.83—272.14—5.04——8063————
MAR-B + l-DeTok#params=208M, Training Protocol=scratch, 400ep2025.071.83—279.6——————————
EDM2-XXLEMA_length_main=0.072024.061.91————————————
SiT-XL/2 + REPA#params=675M2025.072.08—274.6——————————
EDM2-S + Autoguidance (Reduce capacity only)guiding_model_capacity=XS, training_amount=T/16, guidance_weight=1.8, EMA_length_main=0.12, EMA_length_guiding=0.162024.062.13————————————
BA-solverNumber of Function Evaluations (NFE)=102026.012.16—286.38—6.17——8063————
DeCo-XL/16Params=682M, Epochs=340, NFE=100x2, Latency(s)=2.252025.112.22—290—4.67——8060————
EDM2-S + Classifier-free guidanceguidance_weight=1.4, EMA_length_main=0.025, EMA_length_guiding=0.0252024.062.23————————————
AdaGen-MaskGIT-LParams=232M, Steps=32, TFLOPs=3.62026.032.46————————————
CaT-L1#MACs (Per Step)=12G, #Steps=20, #Params=486M2024.042.48——10.64—————————
MaskDiTParams=736M, Steps=792026.032.5————————————
Large-DiT-3B-GParams=3B, Steps=2502026.032.52————————————
EDM2-SEMA_length_main=0.132024.062.56————————————
SiT-XL/2#params=675M2025.072.62—252.2——————————
SiT-XL/2Params=675M + 86M, Epochs=600, NFE=250x2, Latency(s)=11.1, Type=Latent Diffusion Model (VAE)2025.112.62—252.2—4.18——8457————
VDM++Params=2.4B, Steps=2562026.032.65————————————
VDM++Params=2B, Epochs=800, NFE=250x22025.112.65—278.1——————————
DiffiTParams=561M, Steps=2502026.032.67————————————
CaT-L0#MACs (Per Step)=10G, #Steps=20, #Params=377M2024.042.78——14.25—————————
Flow-DPMNumber of Function Evaluations (NFE)=152026.012.79—287.26—7.79——8058————
AdaGen-MaskGIT-LParams=232M, Steps=16, TFLOPs=2.02026.032.82————————————
PixNerd-XL/16Params=700M, Epochs=340, NFE=100x2, Latency(s)=2.472025.112.84—245.6—5.95——8059————
BA-solverNumber of Function Evaluations (NFE)=72026.012.88—300.79—8.35——8061————
EDM2-XSParams=125M, Steps=63, TFLOPs=5.72026.032.91————————————
Simple-DiffusionParams=2B, Steps=5122026.033.02————————————
DiT-XL/2#MACs (Per Step)=525G, #Steps=250, #Params=675M2024.043.04——12.03—————————
DiT-XL/2-GFLOPS=525G, Throughput (A100, B=64, samples/sec)=51, Resolution=512x5122024.113.04————————————
DiT-XL/2#params=675M2025.073.04—240.8——————————
DiT-XLParams=675M, Steps=250, TFLOPs=60.02026.033.04————————————
DiT-XL/2Params=675M + 86M, Epochs=600, NFE=250x2, Latency(s)=11.1, Type=Latent Diffusion Model (VAE)2025.113.04—240.8—5.02——8454————
CFGBackbone=DiT-XL/2, Resolution=512x512, Guidance strategy=Classifier-Free Guidance, CFG scale=1.52025.123.08——————83.853.2—1.219123.68
AsCANFLOPS=224G, Throughput (A100, B=64, samples/sec)=130, Resolution=512x512, sampled cfg=true2024.113.15————————————
AsCANFLOPS=224G, Throughput (A100, B=64, samples/sec)=130, Resolution=512x5122024.113.28————————————
Flow-UniPCNumber of Function Evaluations (NFE)=152026.013.34—272.55—7.99——8057————
SimpleDiffusionParams=2B, Epochs=800, NFE=250x22025.113.54—205——————————
Heun-SolverNumber of Function Evaluations (NFE)=152026.013.58—250.3—7.82——7961————
Euler-SolverNumber of Function Evaluations (NFE)=152026.013.61—266.41—8.57——8057————
DyDiT-XLParams=678M, Steps=100, TFLOPs=75.12026.033.61————————————
ADM-U#MACs (Per Step)=2813G, #Steps=250, #Params=730M2024.043.85——9.96—————————
Flow-DPMNumber of Function Evaluations (NFE)=102026.013.85—304.49—8.83——8055————
RINParams=320M, Steps=1000, TFLOPs=415.02026.033.95————————————
RINParams=320M, NFE=2502025.113.95—210——————————
Token-CriticParams=422M, Steps=36, TFLOPs=34.8, rejection sampling=true2026.034.03————————————
CAN (UViT-S-Deep/4)#MACs (Per Step)=16G, #Steps=50, #Params=185M2024.044.04——23.4—————————
UViT-H/4#MACs (Per Step)=133G, #Steps=50, #Params=501M2024.044.05————————————
U-ViT-H/2FLOPS=546G, Throughput (A100, B=64, samples/sec)=45, Resolution=512x5122024.114.05————————————
U-ViT-HParams=501M, Steps=50, TFLOPs=14.62026.034.05————————————
simple diffusion (U-Net)Resolution=512x512, Model Category=Denoising diffusion models, Backbone=U-Net2023.014.284.3———————————
MaskGITParams=227M, Steps=12, TFLOPs=13.1, rejection sampling=true2026.034.46————————————
simple diffusion (U-ViT 2B)Resolution=512x512, Model Category=Denoising diffusion models, Backbone=U-ViT, Parameters=2B2023.014.533.54———————————
UViT-L/4#MACs (Per Step)=77G, #Steps=50, #Params=287M2024.044.67————————————
U-ViT-L/2FLOPS=340G, Throughput (A100, B=64, samples/sec)=86, Resolution=512x5122024.114.67————————————
DiTResolution=512 x 512, Latency (s)=4.6796, FLOPs=42.0977, Speed=1.00×2026.054.98—186.18—6.56————————
I-ERGBackbone=DiT-XL/2, Resolution=512x512, Guidance strategy=I-ERG, CFG scale=1.52025.125——————85.443.5—1.477124.64
Euler-SolverNumber of Function Evaluations (NFE)=102026.015.08—269.02—9.86——7854————
BA-solverNumber of Function Evaluations (NFE)=52026.015.18—270.26—12.45——7860————
Flow-UniPCNumber of Function Evaluations (NFE)=102026.015.38—290.84—7.23——7950————
Heun-SolverNumber of Function Evaluations (NFE)=102026.015.64—262.29—8.83——7854————
L2C + TCCResolution=512 x 512, Latency (s)=3.9908, FLOPs=33.9235, Speed=1.17×2026.056.03—178.56—6.94————————
L2CResolution=512 x 512, Latency (s)=3.7654, FLOPs=32.7711, Speed=1.24×2026.056.18—178.29—8.15————————
S^2Backbone=DiT-XL/2, Resolution=512x512, Guidance strategy=S2, CFG scale=1.52025.126.2——————83.10.5—1.178124.17
C2FGBackbone=DiT-XL/2, Guidance scale (ω0)=1, λ=ln 2, Sampler=SDE, Resolution=512x512, Samples=10k, Inference timesteps=1002026.036.54—280.9—19.7——8360————
Flow-DPMNumber of Function Evaluations (NFE)=72026.016.59—255.34—10.99——7754————
Token-CriticParams=422M, Steps=36, TFLOPs=7.6, rejection sampling=false2026.036.8————————————
DiT-XL/2Guidance scale (ω)=1.5, Sampler=SDE, Resolution=512x512, Samples=10k, Inference timesteps=1002026.036.81—229.5—20——8262————
MaskGITParams=227M, Steps=12, TFLOPs=3.3, rejection sampling=false2026.037.32————————————
EMAGBackbone=DiT-XL/2, Resolution=512x512, Guidance strategy=Exponential Moving Average Guidance, CFG scale=1.52025.127.59——————86.49.4—1.343125.66
ADM#MACs (Per Step)=1983G, #Steps=250, #Params=559M2024.047.72——23.24—————————
ADM#params=554M2025.077.72—172.7——————————
ADM-GParams=559M, Steps=250, TFLOPs=579.02026.037.72————————————
ADM-GParams=554M, Epochs=400, NFE=250, Latency(s)=21.2, Type=Latent Diffusion Model (VAE)2025.117.72—172.7—6.57——8753————
Flow-DPMNumber of Function Evaluations (NFE)=52026.018.2—256.66—10.67——7647————
Flow-UniPCNumber of Function Evaluations (NFE)=72026.018.43—243.45—9.2——7549————
Heun-SolverNumber of Function Evaluations (NFE)=72026.019.98—214.1—11.17——7354————
APGBackbone=DiT-XL/2, Resolution=512x512, Guidance strategy=Adaptive Projected Guidance, CFG scale=1.52025.1210.24——————84.934.7—1.102125.25
FORA (N = 2) + TCCResolution=512 x 512, Latency (s)=2.6790, FLOPs=21.9778, Speed=1.75×2026.0510.99—139.27—11.06————————
Euler-SolverNumber of Function Evaluations (NFE)=72026.0111.04—204.68—13.17——7352————
CADSBackbone=DiT-XL/2, Resolution=512x512, Guidance strategy=Context-Aware Diffusion Sampling, CFG scale=1.52025.1211.19——————68.871.9—0.785121.54
FluxAE + FT-SEStage II=DiT-L/2, Training steps=400K, Guidance=none2025.0211.63—————203.5——————
No CFGBackbone=DiT-XL/2, Resolution=512x512, Guidance strategy=None2025.1211.78——————7564.6—1.071121.4
SD-VAE-ft-MSEStage II=DiT-XL/2 (orig), Training steps=3M, Guidance=none2025.0212.03————————————
FORA (N = 2)Resolution=512 x 512, Latency (s)=2.5709, FLOPs=21.0647, Speed=1.82×2026.0512.13—136.98—12.02————————
Euler-SolverNumber of Function Evaluations (NFE)=52026.0113.12—203.2—15.2——7144————
FluxAE (vanilla)Stage II=DiT-L/2, Training steps=400K, Guidance=none2025.0213.13—————249.4——————
FluxAE + FTStage II=DiT-L/2, Training steps=400K, Guidance=none2025.0213.69—————267.7——————
SAGBackbone=DiT-XL/2, Resolution=512x512, Guidance strategy=Self-Attention Guidance, CFG scale=1.52025.1218.58——————84.313.4—1.075125.55
SEGBackbone=DiT-XL/2, Resolution=512x512, Guidance strategy=Smoothing-based Enhanced Guidance, CFG scale=1.52025.1222.51——————78.96.9—1.298124.55
Flow-UniPCNumber of Function Evaluations (NFE)=52026.0123.08—127.42—18.23——6248————
Heun-SolverNumber of Function Evaluations (NFE)=52026.0126.11—114—17.97——5953————
VQGANParams=227M, Steps=10242026.0326.52————————————
MDTv2-ST-speed (iter/s)=0.53, GFLOPs=31.462024.1051.16—29.94—8.57————————