yeowon commited on
Commit
3035e25
Β·
verified Β·
1 Parent(s): 070a2a5

Add high-DPI visual benchmark charts (50-block fidelity, TheoremFlow-Omega speedup, Pareto frontier)

Browse files
.gitattributes CHANGED
@@ -37,3 +37,5 @@ assets/MiniMax_H3_Rung_Omega_50blk_Champion.mp4 filter=lfs diff=lfs merge=lfs -t
37
  assets/MiniMax_H3_TheoremFlow_Omega_Champion.mp4 filter=lfs diff=lfs merge=lfs -text
38
  assets/MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav filter=lfs diff=lfs merge=lfs -text
39
  assets/MiniMax_H3_TheoremFlow_Omega_Champion_keyframes.png filter=lfs diff=lfs merge=lfs -text
 
 
 
37
  assets/MiniMax_H3_TheoremFlow_Omega_Champion.mp4 filter=lfs diff=lfs merge=lfs -text
38
  assets/MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav filter=lfs diff=lfs merge=lfs -text
39
  assets/MiniMax_H3_TheoremFlow_Omega_Champion_keyframes.png filter=lfs diff=lfs merge=lfs -text
40
+ assets/minimax_h3_rung_omega_benchmark_hero.png filter=lfs diff=lfs merge=lfs -text
41
+ assets/minimax_h3_pareto_frontier_chart.png filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -48,6 +48,8 @@ tags:
48
 
49
  ## πŸ“Š Benchmark Results
50
 
 
 
51
  ### 1. Full 50-Block End-to-End Velocity Fidelity vs. Official `#1` Recommended `int8_convrot`
52
 
53
  Evaluated on held-out multimodal diffusion trajectories (`Video + Audio + Text + Conditioning Keyframe` across noise levels $\sigma \in [0.08, 0.92]$) against the unquantized `BF16` Teacher (`37.46 GiB`):
@@ -60,6 +62,8 @@ Evaluated on held-out multimodal diffusion trajectories (`Video + Audio + Text +
60
 
61
  ### 2. Multi-Track Pareto Ladder (10-Block Calibration & Held-Out Benchmark)
62
 
 
 
63
  | Track | Trits / Codebook | BPW (Joint-tANS) | Total Size (GiB) | Compression vs. BF16 | Combined Top-1 (↑) | Combined SNR (dB) | Video Cosine (↑) | Audio Cosine (↑) |
64
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
65
  | `baseline_nvfp4_rtn` | NVFP4 | `4.500` | `11.67 GiB` | `3.21x` | `97.53%` | `33.61 dB` | `0.999810` | `0.999676` |
@@ -91,6 +95,8 @@ yeowon/MiniMax-H3-Rung-Omega/
91
  β”‚ β”œβ”€β”€ rung_omega_minimax_h3_50blk_results.json # Full 50-block evaluation metrics vs. int8_convrot
92
  β”‚ └── rung_omega_minimax_h3_results.json # Multi-track Pareto ladder benchmark metrics
93
  └── assets/
 
 
94
  β”œβ”€β”€ MiniMax_H3_TheoremFlow_Omega_Champion.mp4 # 70.11s TheoremFlow-Ξ© generated video + Korean speech
95
  β”œβ”€β”€ MiniMax_H3_Rung_Omega_50blk_Champion.mp4 # Full 20-step generated video + Korean speech
96
  β”œβ”€β”€ MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav # 32kHz stereo audio output
 
48
 
49
  ## πŸ“Š Benchmark Results
50
 
51
+ ![MiniMax-H3-Rung-Omega Benchmark Overview](./assets/minimax_h3_rung_omega_benchmark_hero.png)
52
+
53
  ### 1. Full 50-Block End-to-End Velocity Fidelity vs. Official `#1` Recommended `int8_convrot`
54
 
55
  Evaluated on held-out multimodal diffusion trajectories (`Video + Audio + Text + Conditioning Keyframe` across noise levels $\sigma \in [0.08, 0.92]$) against the unquantized `BF16` Teacher (`37.46 GiB`):
 
62
 
63
  ### 2. Multi-Track Pareto Ladder (10-Block Calibration & Held-Out Benchmark)
64
 
65
+ ![Rung Ξ© Rate-Distortion Pareto Frontier](./assets/minimax_h3_pareto_frontier_chart.png)
66
+
67
  | Track | Trits / Codebook | BPW (Joint-tANS) | Total Size (GiB) | Compression vs. BF16 | Combined Top-1 (↑) | Combined SNR (dB) | Video Cosine (↑) | Audio Cosine (↑) |
68
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
69
  | `baseline_nvfp4_rtn` | NVFP4 | `4.500` | `11.67 GiB` | `3.21x` | `97.53%` | `33.61 dB` | `0.999810` | `0.999676` |
 
95
  β”‚ β”œβ”€β”€ rung_omega_minimax_h3_50blk_results.json # Full 50-block evaluation metrics vs. int8_convrot
96
  β”‚ └── rung_omega_minimax_h3_results.json # Multi-track Pareto ladder benchmark metrics
97
  └── assets/
98
+ β”œβ”€β”€ minimax_h3_rung_omega_benchmark_hero.png # 50-Block Size, Velocity Top-1 & TheoremFlow-Ξ© 8.92x Speedup Chart
99
+ β”œβ”€β”€ minimax_h3_pareto_frontier_chart.png # Rate-Distortion Pareto Frontier vs. NVFP4 & int8_convrot Chart
100
  β”œβ”€β”€ MiniMax_H3_TheoremFlow_Omega_Champion.mp4 # 70.11s TheoremFlow-Ξ© generated video + Korean speech
101
  β”œβ”€β”€ MiniMax_H3_Rung_Omega_50blk_Champion.mp4 # Full 20-step generated video + Korean speech
102
  β”œβ”€β”€ MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav # 32kHz stereo audio output
assets/minimax_h3_pareto_frontier_chart.png ADDED

Git LFS Details

  • SHA256: 2489608c2b91f7420c7805c072999b4aa6a4153c332dacf3acc92243d8ec6150
  • Pointer size: 131 Bytes
  • Size of remote file: 338 kB
assets/minimax_h3_rung_omega_benchmark_hero.png ADDED

Git LFS Details

  • SHA256: ec100616d272bac571eba182dfa0561fc27bbfb3ae499b4185140a3250f9409d
  • Pointer size: 131 Bytes
  • Size of remote file: 340 kB