update heatmaps and other visualizations

arena_elo/elo_rating/elo_analysis.py

@@ -364,6 +364,18 @@ if __name__ == "__main__":
     pretty_print_elo_rating(anony_results["elo_rating_final"])
     print(f"Annoy last update : {anony_results['last_updated_datetime']}")
     print(f"Full last update : {full_results['last_updated_datetime']}")
+    
+    
+    # save heatmap results in the same directory of the cleaned battle file
+    win_fraction_heatmap_file = args.clean_battle_file.replace(".json", "_win_fraction_heatmap.jpg")
+    battle_count_heatmap_file = args.clean_battle_file.replace(".json", "_battle_count_heatmap.jpg")
+    average_win_rate_bar_file = args.clean_battle_file.replace(".json", "_average_win_rate_bar.jpg")
+    bootstrap_elo_rating_file = args.clean_battle_file.replace(".json", "_bootstrap_elo_rating.jpg")
+    anony_results["win_fraction_heatmap"].write_image(win_fraction_heatmap_file)
+    anony_results["battle_count_heatmap"].write_image(battle_count_heatmap_file)
+    anony_results["average_win_rate_bar"].write_image(average_win_rate_bar_file)
+    anony_results["bootstrap_elo_rating"].write_image(bootstrap_elo_rating_file)
+    
 
     last_updated_tstamp = full_results["last_updated_tstamp"]
     cutoff_date = datetime.datetime.fromtimestamp(

arena_elo/results/20240602/elo_results_image_editing.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f987c276a9f91055a204fee98a63f32895c94264e3a4fcc41b84776592763e6f
+size 62362

arena_elo/results/20240602/elo_results_t2i_generation.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:453ed9ef347f96da1353a71fd654e32ed3921d424624a2ec84bade0357e443dc
+size 68024

arena_elo/results/20240602/elo_results_video_generation.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c204c041f0dd9dc27962e14f5b046d931020fe542ef51c265b4dd3e7ce7f3041
+size 59813

arena_elo/results/20240602/image_editing_leaderboard.csv

@@ -0,0 +1,10 @@
+key,Model,Arena Elo rating (anony),Arena Elo rating (full),License,Organization,Link
+MagicBrush,MagicBrush,1111.287883287791,1112.8809479007448,CC-BY-4.0,"The Ohio State University, University of Waterloo",https://osu-nlp-group.github.io/MagicBrush
+InfEdit,InfEdit,1079.0215496678925,1081.8081725884626,CC BY-NC-ND 4.0,"University of Michigan, University of California, Berkeley",https://huggingface.co/spaces/sled-umich/InfEdit
+CosXLEdit,CosXLEdit,1065.6805346660533,1066.704128727434,cosxl-nc-community,Stability AI,https://huggingface.co/spaces/multimodalart/cosxl
+InstructPix2Pix,InstructPix2Pix,1032.7530333195389,1030.5587019503707,"Copyright 2023 Timothy Brooks, Aleksander Holynski, Alexei A. Efros","University of California, Berkeley",https://www.timothybrooks.com/instruct-pix2pix
+PNP,PNP,998.4983696664473,1003.2359999540579,-,Weizmann Institute of Science,https://github.com/MichalGeyer/plug-and-play
+Prompt2prompt,Prompt2prompt,987.9337296502758,989.0974614471744,Apache-2.0,"Google, Tel Aviv University",https://prompt-to-prompt.github.io
+CycleDiffusion,CycleDiffusion,938.9517299322257,932.7970251208932,X11,Carnegie Mellon University,https://github.com/ChenWu98/cycle-diffusion
+SDEdit,SDEdit,928.5595445205557,926.7905094883904,MIT License,Stanford University,https://sde-image-editing.github.io
+Pix2PixZero,Pix2PixZero,857.3136252892202,856.1270528224718,MIT License,"Carnegie Mellon University, Adobe Research",https://pix2pixzero.github.io

arena_elo/results/20240602/t2i_generation_leaderboard.csv

@@ -0,0 +1,12 @@
+key,Model,Arena Elo rating (anony),Arena Elo rating (full),License,Organization,Link
+PlayGround V2.5,PlayGround V2.5,1148.136678409811,1150.447617150882,Playground v2.5 Community License,Playground,https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic
+PlayGround V2,PlayGround V2,1099.2495842883393,1098.3769624420538,Playground v2 Community License,Playground,https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic
+StableCascade,StableCascade,1056.7813216192726,1061.2775980734832,stable-cascade-nc-community (other),Stability AI,https://huggingface.co/stabilityai/stable-cascade
+SDXLLightning,SDXLLightning,1056.3881629701646,1059.8359983006494,openrail++,ByteDance,https://huggingface.co/ByteDance/SDXL-Lightning
+PixArtSigma,PixArtSigma,1050.6588983909805,1049.9019852589274,openrail++,PixArt-alpha,https://fal.ai/models/fal-ai/pixart-sigma
+PixArtAlpha,PixArtAlpha,1050.1258204922178,1038.1819010171778,openrail++,PixArt-alpha,https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS
+SDXL,SDXL,1001.1847782879587,999.8580426291918,openrail++,Stability AI,https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
+SDXLTurbo,SDXLTurbo,935.2347119130764,933.049481437306,sai-nc-community (other),Stability AI,https://huggingface.co/stabilityai/sdxl-turbo
+LCM(v1.5/XL),LCM(v1.5/XL),932.3595672104603,925.8305369502356,openrail++,Latent Consistency,https://fal.ai/models/fal-ai/fast-lcm-diffusion/api
+OpenJourney,OpenJourney,853.094967056503,847.6782124021958,creativeml-openrail-m,PromptHero,https://huggingface.co/prompthero/openjourney
+LCM,LCM,816.7855093612186,829.9051004486075,MIT License,Tsinghua University,https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7

arena_elo/results/20240602/video_generation_leaderboard.csv

@@ -0,0 +1,9 @@
+key,Model,Arena Elo rating (anony),Arena Elo rating (full),License,Organization,Link
+StableVideoDiffusion,StableVideoDiffusion,1148.6551122337614,1159.3538266276412,stable-video-diffusion-nc-community,Stability AI,https://fal.ai/models/fal-ai/fast-svd/text-to-video/api
+T2VTurbo,T2VTurbo,1134.3022495677073,1124.4404105066733,cc-by-nc-4.0,"University of California, Santa Barbara",https://huggingface.co/jiachenli-ucsb/T2V-Turbo-VC2
+AnimateDiff,AnimateDiff,1068.182280658803,1066.112960658009,creativeml-openrail-m,"The Chinese University of Hong Kong, Shanghai AI Lab, Stanford University",https://fal.ai/models/fast-animatediff-t2v
+VideoCrafter2,VideoCrafter2,1062.2218831371833,1063.1081871468969,Apache 2.0,Tencent AI Lab,https://ailab-cvc.github.io/videocrafter2/
+LaVie,LaVie,982.0030996073986,982.0143845325545,Apache 2.0,Shanghai AI Lab,https://github.com/Vchitect/LaVie
+OpenSora,OpenSora,904.4114353398954,904.0996238467394,Apache 2.0,HPC-AI Tech,https://github.com/hpcaitech/Open-Sora
+ModelScope,ModelScope,856.7087803807375,856.3438942029966,cc-by-nc-4.0,Alibaba Group,https://arxiv.org/abs/2308.06571
+AnimateDiffTurbo,AnimateDiffTurbo,843.5151590745159,844.5267124784873,creativeml-openrail-m,"The Chinese University of Hong Kong, Shanghai AI Lab, Stanford University",https://fal.ai/models/fast-animatediff-t2v-turbo

arena_elo/results/latest/clean_battle_image_editing.json

@@ -8950,5 +8950,13 @@
     "judge": "arena_user_10.16.25.191",
     "anony": true,
     "tstamp": 1717229294.5639
+  },
+  {
+    "model_a": "PNP",
+    "model_b": "SDEdit",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717347216.3245
   }
 ]

arena_elo/results/latest/clean_battle_t2i_generation.json

@@ -42166,5 +42166,21 @@
     "judge": "arena_user_10.16.47.67",
     "anony": true,
     "tstamp": 1717282264.3718
+  },
+  {
+    "model_a": "OpenJourney",
+    "model_b": "StableCascade",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717345937.8651
+  },
+  {
+    "model_a": "PlayGround V2.5",
+    "model_b": "StableCascade",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717345942.1509
   }
 ]

arena_elo/results/latest/clean_battle_video_generation.json

@@ -9894,5 +9894,877 @@
     "judge": "arena_user_10.16.27.21",
     "anony": true,
     "tstamp": 1717293775.9663
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717293782.7841
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "VideoCrafter2",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717293783.2738
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "VideoCrafter2",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717293791.6535
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717293798.9666
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "LaVie",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717293806.4462
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717293811.6641
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "ModelScope",
+    "winner": "tie",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717293826.6726
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "T2VTurbo",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717293828.377
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "LaVie",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717293835.7519
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "OpenSora",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717293841.7666
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "ModelScope",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717293848.3812
+  },
+  {
+    "model_a": "AnimateDiff",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717293855.3997
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "T2VTurbo",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717293857.3097
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717293872.4616
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717293901.5198
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "AnimateDiff",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717293932.4272
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717293953.9142
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717293984.306
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "StableVideoDiffusion",
+    "winner": "tie",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294009.905
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "ModelScope",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294030.5896
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "OpenSora",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294056.7294
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "AnimateDiff",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294075.7122
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "AnimateDiff",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294090.8974
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "OpenSora",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294101.2133
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "ModelScope",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294117.938
+  },
+  {
+    "model_a": "AnimateDiff",
+    "model_b": "LaVie",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294150.2901
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "tie",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294174.1211
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "LaVie",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294185.8416
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "OpenSora",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294202.8671
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "VideoCrafter2",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294220.1908
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "ModelScope",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294236.012
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294251.8078
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "OpenSora",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294269.6838
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "OpenSora",
+    "winner": "tie",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294286.4136
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "LaVie",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294312.2032
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294326.2786
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294346.6054
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "VideoCrafter2",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294361.8835
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "ModelScope",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294374.2527
+  },
+  {
+    "model_a": "AnimateDiff",
+    "model_b": "OpenSora",
+    "winner": "tie",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294392.4334
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "AnimateDiff",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294404.4511
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "VideoCrafter2",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294435.7496
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "LaVie",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294453.0766
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "AnimateDiff",
+    "winner": "tie",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294470.3027
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "StableVideoDiffusion",
+    "winner": "tie",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294493.5437
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "OpenSora",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294510.6692
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "AnimateDiff",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294525.3896
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294538.5569
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "VideoCrafter2",
+    "winner": "tie",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294565.9504
+  },
+  {
+    "model_a": "AnimateDiff",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "tie",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294582.533
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "AnimateDiff",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294597.1071
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "VideoCrafter2",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294615.7967
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "OpenSora",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294631.1305
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294644.5529
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "T2VTurbo",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294654.225
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "AnimateDiff",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294668.6136
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "VideoCrafter2",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294682.9829
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294700.8663
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "ModelScope",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294717.4464
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "OpenSora",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294732.9285
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294748.3004
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "AnimateDiff",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294766.0902
+  },
+  {
+    "model_a": "AnimateDiff",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294784.2806
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294796.1612
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "AnimateDiff",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294824.43
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "VideoCrafter2",
+    "winner": "tie",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294851.174
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294866.7995
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "T2VTurbo",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294877.1216
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "ModelScope",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294890.1942
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "VideoCrafter2",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294921.5086
+  },
+  {
+    "model_a": "AnimateDiff",
+    "model_b": "OpenSora",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294934.4353
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "VideoCrafter2",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717294950.2096
+  },
+  {
+    "model_a": "AnimateDiff",
+    "model_b": "LaVie",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717294957.4234
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717294969.0
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "StableVideoDiffusion",
+    "winner": "tie",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717294986.4772
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "VideoCrafter2",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295008.1658
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "ModelScope",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295019.438
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "AnimateDiff",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295038.928
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "OpenSora",
+    "winner": "tie",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295052.3222
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "T2VTurbo",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295070.2534
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717295082.1467
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "LaVie",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295101.6011
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295116.9783
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295126.899
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717295142.3389
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717295154.0226
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295169.8551
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "OpenSora",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717295185.5925
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "StableVideoDiffusion",
+    "winner": "tie",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717295197.1667
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "VideoCrafter2",
+    "winner": "tie",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295210.8272
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295218.0937
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "OpenSora",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295237.2382
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295249.9143
+  },
+  {
+    "model_a": "AnimateDiffTurbo",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295258.4514
+  },
+  {
+    "model_a": "AnimateDiff",
+    "model_b": "AnimateDiffTurbo",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295297.0381
+  },
+  {
+    "model_a": "ModelScope",
+    "model_b": "LaVie",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295309.6869
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "VideoCrafter2",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295320.7567
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "StableVideoDiffusion",
+    "winner": "tie",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717295333.2544
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "VideoCrafter2",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717295349.6526
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "AnimateDiff",
+    "winner": "tie (bothbad)",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295360.2828
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "ModelScope",
+    "winner": "tie",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295370.8103
+  },
+  {
+    "model_a": "StableVideoDiffusion",
+    "model_b": "VideoCrafter2",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295385.1488
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "VideoCrafter2",
+    "winner": "model_a",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295392.7601
+  },
+  {
+    "model_a": "T2VTurbo",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.27.21",
+    "anony": true,
+    "tstamp": 1717295408.8097
+  },
+  {
+    "model_a": "VideoCrafter2",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295422.4892
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "StableVideoDiffusion",
+    "winner": "tie",
+    "judge": "arena_user_10.16.47.67",
+    "anony": true,
+    "tstamp": 1717295436.9749
+  },
+  {
+    "model_a": "LaVie",
+    "model_b": "T2VTurbo",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295449.8035
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "AnimateDiff",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.15.131",
+    "anony": true,
+    "tstamp": 1717295467.5485
+  },
+  {
+    "model_a": "OpenSora",
+    "model_b": "StableVideoDiffusion",
+    "winner": "model_b",
+    "judge": "arena_user_10.16.3.126",
+    "anony": true,
+    "tstamp": 1717345483.6603
   }
 ]

arena_elo/results/latest/elo_results_image_editing.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:77f01bb69a84642304bfbdad304005aee1279b9ea50e7bdb2c40fd8174d94052
-size 62344
+oid sha256:f987c276a9f91055a204fee98a63f32895c94264e3a4fcc41b84776592763e6f
+size 62362

arena_elo/results/latest/elo_results_t2i_generation.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:874b317ce16c1b3c3b0616a5ad5a9ef26b754121a0e241a3e23356f944d00b3b
-size 68005
+oid sha256:453ed9ef347f96da1353a71fd654e32ed3921d424624a2ec84bade0357e443dc
+size 68024

arena_elo/results/latest/elo_results_video_generation.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f0a55a2bcf95eb42d2fc282d1953a7ef9a9403ed71e25b75ce3e8dbd48c6aba4
-size 59749
+oid sha256:c204c041f0dd9dc27962e14f5b046d931020fe542ef51c265b4dd3e7ce7f3041
+size 59813

arena_elo/results/latest/image_editing_leaderboard.csv

@@ -1,10 +1,10 @@
 key,Model,Arena Elo rating (anony),Arena Elo rating (full),License,Organization,Link
-MagicBrush,MagicBrush,1111.2732859076214,1112.8622355116509,CC-BY-4.0,"The Ohio State University, University of Waterloo",https://osu-nlp-group.github.io/MagicBrush
-InfEdit,InfEdit,1079.0224711699423,1081.8108089021207,CC BY-NC-ND 4.0,"University of Michigan, University of California, Berkeley",https://huggingface.co/spaces/sled-umich/InfEdit
-CosXLEdit,CosXLEdit,1065.6784788847476,1066.700916179776,cosxl-nc-community,Stability AI,https://huggingface.co/spaces/multimodalart/cosxl
-InstructPix2Pix,InstructPix2Pix,1032.776893862205,1030.572277544514,"Copyright 2023 Timothy Brooks, Aleksander Holynski, Alexei A. Efros","University of California, Berkeley",https://www.timothybrooks.com/instruct-pix2pix
-PNP,PNP,997.398661562473,1002.1810597126929,-,Weizmann Institute of Science,https://github.com/MichalGeyer/plug-and-play
-Prompt2prompt,Prompt2prompt,987.9308386123184,989.0980692125365,Apache-2.0,"Google, Tel Aviv University",https://prompt-to-prompt.github.io
-CycleDiffusion,CycleDiffusion,938.9136055945952,932.7722089669647,X11,Carnegie Mellon University,https://github.com/ChenWu98/cycle-diffusion
-SDEdit,SDEdit,929.6998472476807,927.8882066556056,MIT License,Stanford University,https://sde-image-editing.github.io
-Pix2PixZero,Pix2PixZero,857.3059171584165,856.1142173141388,MIT License,"Carnegie Mellon University, Adobe Research",https://pix2pixzero.github.io
+MagicBrush,MagicBrush,1111.287883287791,1112.8809479007448,CC-BY-4.0,"The Ohio State University, University of Waterloo",https://osu-nlp-group.github.io/MagicBrush
+InfEdit,InfEdit,1079.0215496678925,1081.8081725884626,CC BY-NC-ND 4.0,"University of Michigan, University of California, Berkeley",https://huggingface.co/spaces/sled-umich/InfEdit
+CosXLEdit,CosXLEdit,1065.6805346660533,1066.704128727434,cosxl-nc-community,Stability AI,https://huggingface.co/spaces/multimodalart/cosxl
+InstructPix2Pix,InstructPix2Pix,1032.7530333195389,1030.5587019503707,"Copyright 2023 Timothy Brooks, Aleksander Holynski, Alexei A. Efros","University of California, Berkeley",https://www.timothybrooks.com/instruct-pix2pix
+PNP,PNP,998.4983696664473,1003.2359999540579,-,Weizmann Institute of Science,https://github.com/MichalGeyer/plug-and-play
+Prompt2prompt,Prompt2prompt,987.9337296502758,989.0974614471744,Apache-2.0,"Google, Tel Aviv University",https://prompt-to-prompt.github.io
+CycleDiffusion,CycleDiffusion,938.9517299322257,932.7970251208932,X11,Carnegie Mellon University,https://github.com/ChenWu98/cycle-diffusion
+SDEdit,SDEdit,928.5595445205557,926.7905094883904,MIT License,Stanford University,https://sde-image-editing.github.io
+Pix2PixZero,Pix2PixZero,857.3136252892202,856.1270528224718,MIT License,"Carnegie Mellon University, Adobe Research",https://pix2pixzero.github.io

arena_elo/results/latest/t2i_generation_leaderboard.csv

@@ -1,12 +1,12 @@
 key,Model,Arena Elo rating (anony),Arena Elo rating (full),License,Organization,Link
-PlayGround V2.5,PlayGround V2.5,1147.6291703802503,1150.0213074790438,Playground v2.5 Community License,Playground,https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic
-PlayGround V2,PlayGround V2,1099.198401481928,1098.3295956022835,Playground v2 Community License,Playground,https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic
-StableCascade,StableCascade,1058.0568819158823,1062.336941809249,stable-cascade-nc-community (other),Stability AI,https://huggingface.co/stabilityai/stable-cascade
-SDXLLightning,SDXLLightning,1056.4503186176216,1059.90290087973,openrail++,ByteDance,https://huggingface.co/ByteDance/SDXL-Lightning
-PixArtSigma,PixArtSigma,1050.6969578706228,1049.9384351522287,openrail++,PixArt-alpha,https://fal.ai/models/fal-ai/pixart-sigma
-PixArtAlpha,PixArtAlpha,1050.0916419067667,1038.1443879907235,openrail++,PixArt-alpha,https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS
-SDXL,SDXL,1001.1475063334567,999.8407416161948,openrail++,Stability AI,https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
-SDXLTurbo,SDXLTurbo,935.1781836808527,932.9973793492295,sai-nc-community (other),Stability AI,https://huggingface.co/stabilityai/sdxl-turbo
-LCM(v1.5/XL),LCM(v1.5/XL),932.3939713973384,925.8504930231434,openrail++,Latent Consistency,https://fal.ai/models/fal-ai/fast-lcm-diffusion/api
-OpenJourney,OpenJourney,852.5184836284428,847.1843181871366,creativeml-openrail-m,PromptHero,https://huggingface.co/prompthero/openjourney
-LCM,LCM,816.6384827868413,829.7832797980595,MIT License,Tsinghua University,https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7
+PlayGround V2.5,PlayGround V2.5,1148.136678409811,1150.447617150882,Playground v2.5 Community License,Playground,https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic
+PlayGround V2,PlayGround V2,1099.2495842883393,1098.3769624420538,Playground v2 Community License,Playground,https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic
+StableCascade,StableCascade,1056.7813216192726,1061.2775980734832,stable-cascade-nc-community (other),Stability AI,https://huggingface.co/stabilityai/stable-cascade
+SDXLLightning,SDXLLightning,1056.3881629701646,1059.8359983006494,openrail++,ByteDance,https://huggingface.co/ByteDance/SDXL-Lightning
+PixArtSigma,PixArtSigma,1050.6588983909805,1049.9019852589274,openrail++,PixArt-alpha,https://fal.ai/models/fal-ai/pixart-sigma
+PixArtAlpha,PixArtAlpha,1050.1258204922178,1038.1819010171778,openrail++,PixArt-alpha,https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS
+SDXL,SDXL,1001.1847782879587,999.8580426291918,openrail++,Stability AI,https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
+SDXLTurbo,SDXLTurbo,935.2347119130764,933.049481437306,sai-nc-community (other),Stability AI,https://huggingface.co/stabilityai/sdxl-turbo
+LCM(v1.5/XL),LCM(v1.5/XL),932.3595672104603,925.8305369502356,openrail++,Latent Consistency,https://fal.ai/models/fal-ai/fast-lcm-diffusion/api
+OpenJourney,OpenJourney,853.094967056503,847.6782124021958,creativeml-openrail-m,PromptHero,https://huggingface.co/prompthero/openjourney
+LCM,LCM,816.7855093612186,829.9051004486075,MIT License,Tsinghua University,https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7

arena_elo/results/latest/video_generation_leaderboard.csv

@@ -1,9 +1,9 @@
 key,Model,Arena Elo rating (anony),Arena Elo rating (full),License,Organization,Link
-StableVideoDiffusion,StableVideoDiffusion,1197.3891415213045,1228.7326412132002,stable-video-diffusion-nc-community,Stability AI,https://fal.ai/models/fal-ai/fast-svd/text-to-video/api
-AnimateDiff,AnimateDiff,1076.7073236982264,1076.2594478514502,creativeml-openrail-m,"The Chinese University of Hong Kong, Shanghai AI Lab, Stanford University",https://fal.ai/models/fast-animatediff-t2v
-T2VTurbo,T2VTurbo,1075.1283325084098,1034.2104834546742,cc-by-nc-4.0,"University of California, Santa Barbara",https://huggingface.co/jiachenli-ucsb/T2V-Turbo-VC2
-VideoCrafter2,VideoCrafter2,1066.1439972125954,1068.6824957279891,Apache 2.0,Tencent AI Lab,https://ailab-cvc.github.io/videocrafter2/
-LaVie,LaVie,975.5835007290805,977.4242179063297,Apache 2.0,Shanghai AI Lab,https://github.com/Vchitect/LaVie
-OpenSora,OpenSora,907.78356567307,909.0903616850528,Apache 2.0,HPC-AI Tech,https://github.com/hpcaitech/Open-Sora
-ModelScope,ModelScope,856.5801154935552,858.0685226371207,cc-by-nc-4.0,Alibaba Group,https://arxiv.org/abs/2308.06571
-AnimateDiffTurbo,AnimateDiffTurbo,844.6840231637599,847.531829524186,creativeml-openrail-m,"The Chinese University of Hong Kong, Shanghai AI Lab, Stanford University",https://fal.ai/models/fast-animatediff-t2v-turbo
+StableVideoDiffusion,StableVideoDiffusion,1148.6551122337614,1159.3538266276412,stable-video-diffusion-nc-community,Stability AI,https://fal.ai/models/fal-ai/fast-svd/text-to-video/api
+T2VTurbo,T2VTurbo,1134.3022495677073,1124.4404105066733,cc-by-nc-4.0,"University of California, Santa Barbara",https://huggingface.co/jiachenli-ucsb/T2V-Turbo-VC2
+AnimateDiff,AnimateDiff,1068.182280658803,1066.112960658009,creativeml-openrail-m,"The Chinese University of Hong Kong, Shanghai AI Lab, Stanford University",https://fal.ai/models/fast-animatediff-t2v
+VideoCrafter2,VideoCrafter2,1062.2218831371833,1063.1081871468969,Apache 2.0,Tencent AI Lab,https://ailab-cvc.github.io/videocrafter2/
+LaVie,LaVie,982.0030996073986,982.0143845325545,Apache 2.0,Shanghai AI Lab,https://github.com/Vchitect/LaVie
+OpenSora,OpenSora,904.4114353398954,904.0996238467394,Apache 2.0,HPC-AI Tech,https://github.com/hpcaitech/Open-Sora
+ModelScope,ModelScope,856.7087803807375,856.3438942029966,cc-by-nc-4.0,Alibaba Group,https://arxiv.org/abs/2308.06571
+AnimateDiffTurbo,AnimateDiffTurbo,843.5151590745159,844.5267124784873,creativeml-openrail-m,"The Chinese University of Hong Kong, Shanghai AI Lab, Stanford University",https://fal.ai/models/fast-animatediff-t2v-turbo

requirements.txt

@@ -68,7 +68,4 @@ tensorboard
 timm
 wandb
 pandarallel
-
-
-
-
+kaleido

serve/leaderboard.py

@@ -192,7 +192,7 @@ Contribute your vote 🗳️ at [vision-arena](https://huggingface.co/spaces/Wil
 """
     return leaderboard_md
 
-def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=False):
+def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=True):
     if elo_results_file is None:  # Do live update
         md = "Loading ..."
         p1 = p2 = p3 = p4 = None
@@ -265,12 +265,41 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
                     height=700,
                     wrap=True,
                 )
-        if not show_plot:
-            gr.Markdown(
-                """ ## We are still collecting more votes on more models. The ranking will be updated very fruquently. Please stay tuned! 
-                """,
-                elem_id="leaderboard_markdown",
-            )
+                
+        gr.Markdown(
+            """ ## We are still collecting more votes on more models. The ranking will be updated very fruquently. Please stay tuned! 
+            """,
+            elem_id="leaderboard_markdown",
+        )
+        
+        if show_plot:
+            win_fraction_heatmap = anony_elo_results["win_fraction_heatmap"]
+            battle_count_heatmap = anony_elo_results["battle_count_heatmap"]
+            bootstrap_elo_rating = anony_elo_results["bootstrap_elo_rating"]
+            average_win_rate_bar = anony_elo_results["average_win_rate_bar"]
+            with gr.Row():
+                with gr.Column():
+                    gr.Markdown(
+                        "#### Figure 1: Fraction of Model A Wins for All Non-tied A vs. B Battles"
+                    )
+                    plot_1 = gr.Plot(win_fraction_heatmap, show_label=False)
+                with gr.Column():
+                    gr.Markdown(
+                        "#### Figure 2: Battle Count for Each Combination of Models (without Ties)"
+                    )
+                    plot_2 = gr.Plot(battle_count_heatmap, show_label=False)
+            with gr.Row():
+                with gr.Column():
+                    gr.Markdown(
+                        "#### Figure 3: Bootstrap of Elo Estimates (1000 Rounds of Random Sampling)"
+                    )
+                    plot_3 = gr.Plot(bootstrap_elo_rating, show_label=False)
+                with gr.Column():
+                    gr.Markdown(
+                        "#### Figure 4: Average Win Rate Against All Other Models (Assuming Uniform Sampling and No Ties)"
+                    )
+                    plot_4 = gr.Plot(average_win_rate_bar, show_label=False)
+    
     else:
         pass