List your company

Image Generation on ImageNet-1K 256x256 (val)

478.35Inception ScoreFP
125.218216.8965308.575400.2535Jun 11, 2024Oct 7, 2024Feb 2, 2025May 31, 2025Sep 26, 2025Jan 22, 2026May 20, 2026

Evaluation results

MethodLinks
FPW/A=32/32, cfg=4.0, step=502025.09478.35———92.98———2,575.4114.52
VAR-d30-retokenizer=VQ, type=VAR, #params=2.0B, rejection sampling=true, generation resolution=2562024.11350.21.73——8260————
RobuQW/A=1.58/4, cfg=4.0, step=502025.09349.22———91.86———194.7510.07
RobuQW/A=1.58/3, cfg=4.0, step=502025.09344.29———90.83———194.758.34
BaselineW/A=1.58/4, cfg=4.0, step=502025.09342.07———90.92———194.7510.07
MaskBittokenizer=VQ, type=Mask., #params=305M, rejection sampling=false, generation resolution=2562024.11328.61.52————————
RAR-XXLtokenizer=VQ, type=AR, #params=1.5B, rejection sampling=false, generation resolution=2562024.113261.48——8063————
RQTran.-reType=AR, #Para.=3.8B2026.03323.73.8————————
VAR-d30tokenizer=VQ, type=VAR, #params=2.0B, rejection sampling=false, generation resolution=2562024.11323.11.92——8259————
MAGVIT-v2tokenizer=LFQ, guidance=true, Parameters=307M, Sampling steps=64, Throughput=1.12024.06319.41.78————————
MAGVIT-v2tokenizer=VQ, type=Mask., #params=307M, rejection sampling=false, generation resolution=2562024.11319.41.78————————
MDTv2-XL/2tokenizer=VAE, type=Diff., #params=676M, rejection sampling=false, generation resolution=2562024.11314.71.58——7965————
VAR-d24#Steps=10, Speedup=1.7x, Latency=2.07s, Throughput=30.92it/s, #Param=1.0B, Memory=25093MB2024.11311———82592.11———
RAR-XLtokenizer=VQ, type=AR, #params=955M, rejection sampling=false, generation resolution=2562024.11306.91.5——8062————
MAR-Htokenizer=VAE, type=MAR, #params=943M, rejection sampling=false, generation resolution=2562024.11303.71.55——8162————
VAR-CoDeN=7, #Steps=7+3, Speedup=2.3x, Latency=1.60s, Throughput=40.00it/s, #Param=2.0+0.3B, Memory=19943MB2024.11303———82592.11———
VAR-d20Type=VAR, #Para.=600M2026.03302.62.57——8356————
VAR-CoDeN=8, #Steps=8+2, Speedup=1.7x, Latency=2.11s, Throughput=30.33it/s, #Param=2.0+0.3B, Memory=21019MB2024.11302———81601.98———
VAR-d30#Steps=10, Speedup=1.0x, Latency=3.62s, Throughput=17.71it/s, #Param=2.0B, Memory=39228MB2024.11301———81591.95———
VAR-d20#Steps=10, Speedup=2.8x, Latency=1.29s, Throughput=49.62it/s, #Param=600M, Memory=17814MB2024.11301———83562.61———
RAR-Ltokenizer=VQ, type=AR, #params=461M, rejection sampling=false, generation resolution=2562024.11299.51.7——8160————
VAR-CoDeN=6, #Steps=6+4, Speedup=2.9x, Latency=1.27s, Throughput=50.39it/s, #Param=2.0+0.3B, Memory=19943MB2024.11297———82582.27———
MAR-Ltokenizer=VAE, type=MAR, #params=479M, rejection sampling=false, generation resolution=2562024.112961.78——8160————
VAR-CoDeN=9, #Steps=9+1, Speedup=1.2x, Latency=2.97s, Throughput=21.54it/s, #Param=2.0+0.3B, Memory=28803MB2024.11296———80611.94———
ImageFolderType=VAR, #Para.=362M2026.032952.6——7563————
RAR-Btokenizer=VQ, type=AR, #params=261M, rejection sampling=false, generation resolution=2562024.11290.51.95——8258————
DiMR-XL/2Rtokenizer=VAE, type=Diff., #params=505M, rejection sampling=false, generation resolution=2562024.112891.7——7963————
Static c4Iters.=600K, Params=675M, Sampler=SDE, Guidance=CFG (w=1.5), Chunk size=42026.05287———84572.234.49——
U-DiT-LIters.=250K, Params=810M, Sampler=SDE, Guidance=CFG (w=1.5)2026.05286.6———865234.4——
Dynamic c4Iters.=500K, Params=751M, Sampler=SDE, Guidance=CFG (w=1.5), Chunk size=42026.05284.8———83572.174.49——
MAR-B#Steps=100, Speedup=0.1x, Latency=29.80s, Throughput=2.15it/s, #Param=208M, Memory=8725MB2024.11282———82572.31———
MaskGIT-UViT-Ltokenizer=TiTok-S-128, guidance=true, Parameters=287M, Sampling steps=64, Throughput=7.82024.06281.81.97————————
TiTok-S-128tokenizer=VQ, type=Mask., #params=287M, rejection sampling=false, generation resolution=2562024.11281.81.97————————
MAR-Btokenizer=VAE, type=MAR, #params=208M, rejection sampling=false, generation resolution=2562024.11281.72.31——8257————
GPT2-retokenizer=VQ, type=AR, #params=1.4B, rejection sampling=true, generation resolution=2562024.11280.35.2————————
VQGAN-reType=AR, #Para.=1.4B2026.03280.35.2————————
MaskGIT-UViT-Ltokenizer=TiTok-S-128, guidance=true, Parameters=287M, Sampling steps=8, Throughput=53.32024.06278.72.5————————
DiT-XL/2tokenizer=VAE, guidance=true, Parameters=675M, Sampling steps=250, Throughput=0.62024.06278.22.27————————
DiT-XL/2tokenizer=VAE, type=Diff., #params=675M, rejection sampling=false, generation resolution=2562024.11278.22.27——8357————
DiT-XL/2Type=Diffusion, #Para.=675M2026.03278.22.27——8357————
DiTIters.=1.75M, Params=675M, Sampler=SDE, Guidance=CFG (w=1.5)2026.05278.2———83572.27———
VAR-d16Type=VAR, #Para.=310M2026.03274.43.3——8451————
RobuQW/A=1.58/2, cfg=4.0, step=502025.09273.58———87.51———194.756.61
Static c4Iters.=600K, Params=675M, Sampler=ODE, Guidance=CFG (w=1.5), Chunk size=42026.05272.9———83612.084.42——
Open-MAGVIT2-XLtokenizer=VQ, type=AR, #params=1.5B, rejection sampling=false, generation resolution=2562024.11271.82.33——8454————
Open-MAGVIT2-Ltokenizer=VQ, type=AR, #params=804M, rejection sampling=false, generation resolution=2562024.11271.72.51——8454————
Open-MAGVIT2-AR-LType=AR, #Para.=804M2026.03271.72.51——8454————
SiT-XLtokenizer=VAE, type=Diff., #params=675M, rejection sampling=false, generation resolution=2562024.11270.32.06——8259————
SiTIters.=1.75M, Params=675M, Sampler=SDE, Guidance=CFG (w=1.5)2026.05270.3———82592.064.49——
Dynamic c4Iters.=500K, Params=751M, Sampler=ODE, Guidance=CFG (w=1.5), Chunk size=42026.05270.1———82602.054.39——
UViT-H/2tokenizer=VAE, guidance=true, Parameters=501M, Sampling steps=50, Throughput=0.62024.06263.92.29————————
UViT-H/2tokenizer=VAE, type=Diff., #params=501M, rejection sampling=false, generation resolution=2562024.11263.92.29——8257————
U-ViT-H/2Iters.=500K, Params=585M, Sampler=ODE, Guidance=CFG (w=1.5)2026.05263.9———82572.295.68——
LlamaGen-3B-384tokenizer=VQ, type=AR, #params=3.1B, rejection sampling=false, generation resolution=3842024.11263.32.18——8158————
Open-MAGVIT2-Btokenizer=VQ, type=AR, #params=343M, rejection sampling=false, generation resolution=2562024.11258.33.08——8551————
Open-MAGVIT2-AR-BType=AR, #Para.=343M2026.03258.263.08——8551————
SiTIters.=1.75M, Params=675M, Sampler=ODE, Guidance=CFG (w=1.5)2026.05258.1———81602.154.6——
AiM-XL#Steps=256, Speedup=0.4x, Latency=9.32s, Throughput=6.87it/s, #Param=763M, Memory=20983MB2024.11257———82572.56———
LlamaGen-L-384tokenizer=VQ, type=AR, #params=343M, rejection sampling=false, generation resolution=3842024.11256.13.07——8352————
SiT-PlusIters.=1M, Params=752M, Sampler=SDE, Guidance=CFG (w=1.5)2026.05254.1———82572.344.56——
LlamaGen-XXL#Steps=384, Speedup=<0.1x, Latency=73.97s, Throughput=0.87it/s, #Param=1.4B, Memory=42632MB2024.11254———80592.34———
LlamaGen-XXL-384tokenizer=VQ, type=AR, #params=1.4B, rejection sampling=false, generation resolution=3842024.11253.92.34——8059————
LlamaGen-XXLtokenizer=VQ, type=AR, #params=1.4B, rejection sampling=false, generation resolution=2562024.11253.63.09——8353————
LlamaGen-Ltokenizer=VQ, type=AR, #params=343M, rejection sampling=false, generation resolution=2562024.11248.33.8——8351————
LlamaGen-LType=AR, #Para.=343M2026.03248.283.8——8351————
LDM-4tokenizer=VAE, guidance=true, Parameters=400M, Sampling steps=250, Throughput=0.42024.06247.73.6————————
LDM-4tokenizer=VAE, type=Diff., #params=400M, rejection sampling=false, generation resolution=2562024.11247.73.6——8748————
LDM-4Type=Diffusion, #Para.=400M2026.03247.73.6————————
SiT-PlusIters.=1M, Params=752M, Sampler=ODE, Guidance=CFG (w=1.5)2026.05244.6———80582.364.4——
LlamaGen-XL-384tokenizer=VQ, type=AR, #params=775M, rejection sampling=false, generation resolution=3842024.11244.12.62——8057————
DiTIters.=1.75M, Params=675M, Sampler=ODE, Guidance=CFG (w=1.5)2026.05239.2———80592.254.29——
DiT-XL/2#Steps=50, Speedup=0.2x, Latency=19.20s, Throughput=3.33it/s, #Param=675M, Memory=11369MB2024.11239———80602.26———
LlamaGen-L + ProVQ tokenizerType=AR, #Para.=343M, tokenizer size=16 x 162026.03235.513.15——8254————
VIM-L-retokenizer=VQ, type=AR, #params=1.7B, rejection sampling=true, generation resolution=2562024.11227.43.04————————
ViT-VQGAN-reType=AR, #Para.=1.7B2026.03227.43.04————————
LlamaGen-XLtokenizer=VQ, type=AR, #params=775M, rejection sampling=false, generation resolution=2562024.11227.13.39——8154————
LlamaGen-3Btokenizer=VQ, type=AR, #params=3.1B, rejection sampling=false, generation resolution=2562024.11222.33.05——8058————
UViT-L/2tokenizer=VAE, guidance=true, Parameters=287M, Sampling steps=50, Throughput=1.12024.06219.93.4————————
UViT-L/2tokenizer=VAE, type=Diff., #params=287M, rejection sampling=false, generation resolution=2562024.11219.93.4——8352————
MaskGIT-ViTtokenizer=TiTok-B-64, guidance=true, Parameters=177M, Sampling steps=8, Throughput=89.82024.06214.72.48————————
LDM-8tokenizer=Taming-VQGAN, guidance=true, Parameters=258M, Sampling steps=2002024.06209.57.76————————
LDM-8tokenizer=VQ, type=Diff., #params=258M, rejection sampling=false, generation resolution=2562024.11209.57.76——8435————
MAGVIT-v2tokenizer=LFQ, guidance=false, Parameters=307M, Sampling steps=64, Throughput=1.12024.06200.53.65————————
MaskGIT-ViTtokenizer=TiTok-L-32, guidance=true, Parameters=177M, Sampling steps=8, Throughput=101.62024.06199.82.77————————
Bitnetv2W/A=1.58/4, cfg=4.0, step=502025.09196.78———83.7———148.138.04
LlamaGen-BType=AR, #Para.=111M2026.03193.615.46——8345————
MaskGIT-ViTtokenizer=TiTok-B-64, guidance=false, Parameters=177M, Sampling steps=8, Throughput=89.82024.06192.53.08————————
LlamaGen-B + ProVQ tokenizerType=AR, #Para.=111M, tokenizer size=16 x 162026.03190.34.99——8446————
MaskGIT-ViTtokenizer=MaskGIT-VQGAN, guidance=false, Parameters=177M, Sampling steps=8, Throughput=50.52024.06182.16.18————————
MaskGITtokenizer=VQ, type=Mask., #params=177M, rejection sampling=false, generation resolution=2562024.11182.16.18————————
VIM-Largetokenizer=ViT-VQGAN, guidance=false, Parameters=1.7B, Sampling steps=1024, Throughput=0.32024.06175.14.17————————
VIM-Ltokenizer=VQ, type=AR, #params=1.7B, rejection sampling=false, generation resolution=2562024.11175.14.17————————
ViT-VQGANType=AR, #Para.=1.7B2026.03175.14.17————————
MaskGIT-ViTtokenizer=TiTok-L-32, guidance=false, Parameters=177M, Sampling steps=8, Throughput=101.62024.061733.15————————
MaskGIT-UViT-Ltokenizer=TiTok-S-128, guidance=false, Parameters=287M, Sampling steps=64, Throughput=7.82024.06168.24.44————————
DiT-L/2tokenizer=VAE, type=Diff., #params=458M, rejection sampling=false, generation resolution=2562024.11167.25.02——7557————
MaskGIT-UViT-Ltokenizer=TiTok-S-128, guidance=false, Parameters=287M, Sampling steps=8, Throughput=53.32024.06166.74.61————————
CDMType=Diffusion2026.03158.74.88————————
UViT-H/2tokenizer=VAE, guidance=false, Parameters=501M, Sampling steps=50, Throughput=0.62024.06142.56.6————————
Static c4Iters.=600K, Params=675M, Sampler=SDE, Guidance=None, Chunk size=42026.05138.8———70676.925.27——