qgfvadfuvads/I2V-pairs-m1xm3

I2V Pairs (m1xm3)

Paired image-to-video generations for preference / DPO research. Each entry contains two videos generated from the same source image (first frame of source_video) and same caption, but with two different sampling configs (method m1: CFG=7, steps=15 vs method m3: CFG=3, steps=50) on Wan2.2-TI2V-5B @ 720p.

Stats

• 1284 clean pairs (after filtering for unreadable / static / first-frame-drift)
• Source videos: OpenVid (parts 112-113), 5s clips at 720p
• Generated: 2026-04-24

Files

• metadata.json — list of {id, caption, video_a, video_b, source_video}
• videos.tar — extract to videos/ (2*N pair mp4s)
• source_videos.tar — extract to source_videos/ (N source mp4s)

Quick start

tar xf videos.tar
tar xf source_videos.tar
python -c "import json; d=json.load(open('metadata.json')); print(len(d), d[0])"

Filtering thresholds

Pairs were rejected if any of:

• mp4_unreadable / too_few_frames
• static: mean inter-frame |diff| < 1.5 (frozen / I2V failed)
• first_frame_drift: PSNR(first frame, source first frame) < 18 dB (anchor broken)
• pair_identical: PSNR(pair_0, pair_1) > 35 dB (no real difference)

Per-pair raw metrics (including rejected ones) are kept separately as I2V_pairs_m1xm3_diagnostics.json in the source repo, not shipped here.

License

Source clips inherit from OpenVid-1M (CC-BY 4.0). Generated outputs released CC-BY-NC-4.0.