2026-08-31 15:57:13 +08:00

989 B

MiniMax-H3 paired SSIM: base vs Cache-DiT vs Larry LoRA

Reference: /data/wxy/results/minimax_h3_mixed_matrix/balanced-tp4-tp2-20steps-5s-20260822-175030

Metric: FFmpeg-decoded YUV420 SSIM All, paired by request_id after exact prompt/seed/task/resolution/duration/aspect-ratio checks. Base self-comparison is 1.0.

Scheme Videos Mean Median Video P10 Worst video >= 0.90
base 64 1.000000 1.000000 1.000000 1.000000 64/64
cache_dit 64 0.771867 0.773992 0.624387 0.458895 14/64
lora 64 0.706152 0.707145 0.536034 0.473817 1/64

By task

  • cache_dit / fl2va: mean=0.780862, >=0.90=9/32
  • cache_dit / ref2va: mean=0.762872, >=0.90=5/32
  • lora / fl2va: mean=0.710248, >=0.90=0/32
  • lora / ref2va: mean=0.702055, >=0.90=1/32

By resolution

  • cache_dit: 480p=0.709407, 720p=0.757765, 768p=0.765560, 1080p=0.854737
  • lora: 480p=0.654367, 720p=0.686297, 768p=0.696015, 1080p=0.787927