BlueSourceJY's picture
ShotStream priority 1-200
622e845 verified
|
Raw History Blame Contribute Delete
4.42 kB
metadata
license: other
pretty_name: ERCBench ShotStream (priority 1-200)
tags:
  - video-generation
  - multi-shot
  - text-to-video

ERCBench-ShotStream-400 — ShotStream outputs, priority 1–200

Multi-shot videos generated with ShotStream (KlingAIResearch, base model Wan2.1-T2V-1.3B) for the first 200 episodes (priority 1–200) of the UEC-Bench / ERCBench StoryMem prompt set. This release has 200 episodes and 1,000 shots: 120 four-shot, 60 six-shot and 20 eight-shot episodes.

Contents

episodes/<episode_id>/episode.mp4      # one continuous multi-shot video per episode
episodes/<episode_id>/metadata.json    # prompts + sha256, frame ranges, cut frames, seed, video sha256, timing, GPU
inputs/shard_{0..3}.csv                # ShotStream input CSVs (one per GPU), smoke.csv
inputs/json/<episode_id>.json          # {"global_caption": "", "shot1": ..., "shotN": ...}
runs/run_manifest.json                 # versions, revisions, attention backend, GPU, start/end times (Dubai, UTC+4)
runs/inference_episode.diff            # the only changes made to Inference_Causal.py
runs/timing.csv                        # per-episode generation time
runs/logs/                             # raw generation logs, GPU memory samples
runs/scripts/                          # verification / input-building / finalize scripts, pip freeze

Prompt source

  • Dataset: BlueSourceJY/ERCBench-StoryMem-400, revision c96eccaf5f9de67181aa6bdcd22fae85252d5821, file storymem_inputs_400.zip (sha256 50ac67c03afc261a0228fc894596832676f4126ebf1a0edc5bf2628decc71c2f).
  • All 2,045 files were checked against storymem_inputs/sha256.json.
  • Episodes come from storymem_inputs/priority_200.ids.txt. The file order is the priority order.
  • Shot k of an episode is shot_mapping.json → shots[k-1], which gives (storymem_scene_num, storymem_shot_num). The prompt is story_script.json → the scene with that scene_num → video_prompts[storymem_shot_num - 1].
  • Each prompt was verified against prompt_sha256 (1,000/1,000 matched). Prompts are used verbatim: they are not rewritten, translated or expanded. Only video_prompts are used. global_caption is the empty string. ShotStream's data loader adds the shot{i}: prefix itself.

Generation settings

  • Code: ShotStream commit c6a11719ee8ac203ee20afc27056446fe25cef86.

  • Weights: Wan-AI/Wan2.1-T2V-1.3B @ 37ec512624d61f7aa208f7ea8140a131f93afc9a and KlingTeam/ShotStream (shotstream_merged.pt, shotstream.yaml, default_config.yaml) @ 225806baab3c54e94351474e976948eb5c61b6fe.

  • The configs are unmodified (4 denoising steps [1000, 740, 500, 260], max_context_frames: 6, multi_caption: True), and so are the resolution and per-shot frame count.

  • 832×480, 16 fps, 81 frames per shot, H.264 / yuv420p, written by torchvision.io.write_video.

  • Seed 0, re-set with set_seed(0) before every episode, so the results do not depend on how episodes were sharded.

  • Attention backend: FlashAttention-2 (flash_attn 2.8.3.post2). torch 2.8.0+cu128.

  • Hardware: 4× NVIDIA H200, one independent process per GPU (no torchrun).

  • Inference script: Inference_Causal.py was copied to inference_episode.py with three changes, listed in runs/inference_episode.diff:

    1. output named <episode_id>.mp4;
    2. set_seed(seed) before each episode;
    3. skip an episode whose output already exists with the correct frame count.

    Model loading, pipeline.inference and write_video(fps=16) are unchanged.

  • Shots are generated autoregressively within an episode. Each new shot is conditioned on 6 context frames sampled from the previously generated shots, plus the captions of all shots so far. Shots are not generated independently.

Cut positions

An N-shot episode has 81 × N frames. Shot k (1-based) occupies frames [81(k-1), 81k), so the cuts are at frames 81, 162, …, 81(N-1), which are stored in metadata.json as shot_frame_ranges and cut_frames. Frame-difference analysis confirmed that in all 200 videos the N−1 largest inter-frame changes fall exactly on these cut frames.

Validation

All 200 videos are 832×480 at 16 fps with exactly 81 × shot_count frames, and all decode fully with ffmpeg. There were no failed episodes and no reruns.