sskj-h3/vbench-score/vbench-lora/README.md

MiniMax-H3 Larry v4-600 LoRA VBench-v1

This run is a one-variable comparison against ../vbench-base/results/h3-vbench-base-dense-tp2x4-20260826-run1.

• Same official VBench-v1 metadata and all 16 dimensions
• Same 944 unique prompts, one fixed-seed sample per prompt
• Same seed 1101, 768P, 16:9, 5 seconds, flow shifts 12.0/3.0
• Same four TP2 SGLang services and deterministic prompt balancing (236 each)
• Same official static filter before temporal-flickering scoring
• Changed only the serving environment/LoRA and sampler schedule: SGLang 0.5.18, Larry v4-600 EMA, 9 grid points = 8 denoiser evaluations

Larry v4-600 is intended for FL2VA, which is also the task used by VBench text-to-video generation here.

The launcher is resumable: completed videos recorded in generation/results.jsonl are reused after a safe restart.