ssaroya/atom-1-physics-iq-verified-run01
Atom 1 — Physics-IQ Verified i2v submission (4 runs) Moving Atoms · Physics-IQ Verified · Image-to-Video 4-run aggregate: 47.06 ± 1.32 (4 independent runs, seeds 42/1042/2042/3042). Individual run means: 45.44 / 47.27 / 46.88 / 48.64. System frozen across runs (LoRA weights, planner labels, prompts byte-identical); only the diffusion seed varies. Current i2v board leader: MiniMax H3 at 39.8 ± 0.3 (2026-08-24) — Atom 1's own base checkpoint. Atom 1 is MiniMaxAI/MiniMax-H3 (FL2VA… See the full description on the dataset page: https://huggingface.co/datasets/ssaroya/atom-1-physics-iq-verified-run01.
Atom 1 — Physics-IQ Verified i2v submission (4 runs)
Moving Atoms · Physics-IQ Verified · Image-to-Video
4-run aggregate: 47.06 ± 1.32 (4 independent runs, seeds 42/1042/2042/3042). Individual run means: 45.44 / 47.27 / 46.88 / 48.64. System frozen across runs (LoRA weights, planner labels, prompts byte-identical); only the diffusion seed varies. Current i2v board leader: MiniMax H3 at 39.8 ± 0.3 (2026-08-24) — Atom 1's own base checkpoint.
Atom 1 is MiniMaxAI/MiniMax-H3 (FL2VA bf16 open checkpoint) post-trained with a Moving Atoms physics LoRA (WISA-80K, rank-16 attn) + the Atom Harness: a planner that once assigned each scenario a motion class (settling/continuous/cascade/quiescent) from the switch frame and official description alone; the templater then deterministically writes the class's timed clause into the prompt. One sample per case, no selection.
Contents
run_01/— 198 MP4s, seed 42 (original single-run submission)run_02/— 198 MP4s, seed 1042run_03/— 198 MP4s, seed 2042run_04/— 198 MP4s, seed 3042scores/final_run_stats.json— per-run means + per-case scores for runs 02–04descriptions.csv— the exact per-case prompts used for generationsubmission.yaml— the official submission cardTECH-REPORT.md— the technical report (LoRA, harness, generation, scoring)
All videos machine-generated by Atom 1. 198 per run, exactly 5.000 s each, 1344×768, 24 fps, benchmark naming (0001_…0198_).
Config
DiffSynth-Studio, self-hosted · official reference config (1344×768, 124f @ 24 fps, flowshift 12/3, CFG-distilled) at `numinference_steps=20` · untiled VAE decode · single pass · ~5.5 GPU-min/clip on H100.
Scoring self-reported from a re-implementation validated against the official pipeline at 0.02-point agreement; the maintainers recompute all results independently.
