BlueSourceJY's picture
LongLive priority 1-200
3002d57 verified
|
Raw History Blame Contribute Delete
5.86 kB
---
license: other
pretty_name: ERCBench LongLive (priority 1-200)
tags:
- video-generation
- multi-shot
- text-to-video
---
# ERCBench-LongLive-400 — LongLive outputs, priority 1–200
Multi-prompt videos generated with **LongLive** (NVlabs, LongLive-1.3B, 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. The prompts and the episode list are the same
as in [`BlueSourceJY/ERCBench-ShotStream-400`](https://huggingface.co/datasets/BlueSourceJY/ERCBench-ShotStream-400).
> **Shot boundaries are prompt switches, not hard cuts.** LongLive is an interactive long-video model: it generates
> one continuous video and switches the text prompt during generation (KV re-cache). Each shot here is 21 latent
> frames driven by that shot's prompt, and the transition between shots is LongLive's smooth prompt switch, not a
> camera cut. `shot_frame_ranges` record where each prompt takes over.
## Contents
```
episodes/<episode_id>/episode.mp4 # one continuous video per episode
episodes/<episode_id>/metadata.json # prompts + sha256 + UMT5 token counts, frame ranges, switch frames, seed, video sha256, timing
inputs/prompts_{4,6,8}shot.jsonl # LongLive input files: {"prompts": [Shot1, ..., ShotN], "episode_id": ...}, priority order
inputs/longlive_{4,6,8}shot.yaml # the official interactive config with only paths, frame count and switch indices changed
runs/run_manifest.json # versions, revisions, attention backend, GPU, start/end times (Dubai, UTC+4)
runs/interactive_episode.diff # the only changes made to interactive_inference.py
runs/timing.csv # per-episode generation time
runs/logs/ # generation log, tokenizer log (every string sent to the text encoder), GPU memory
runs/scripts/ # input-building, tokenizer-logging, finalize and packaging scripts, pip freeze
```
## Prompt source
- Dataset: [`BlueSourceJY/ERCBench-StoryMem-400`](https://huggingface.co/datasets/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]`.
`make_inputs.py` re-derives every prompt from these source files and checks it against `prompt_sha256`.
- Prompts are used **verbatim**: not rewritten, translated or expanded, and no global caption or prefix is added.
- Every string sent to LongLive's own UMT5 tokenizer was logged (`runs/logs/tokens.jsonl`, via `tok_hook_run.py`,
log-only). For all 1,000 shots, the string equals the dataset prompt (same sha256), shot k uses Shot k's prompt, and
the untruncated token count is at most 512 (no truncation). Per-shot counts are in each `metadata.json`.
## Generation settings
- Code: [LongLive](https://github.com/NVlabs/LongLive) commit `31c70d9071cd7851d82939806dc9fac645fcabe5`, `LongLive1.0/`.
- Weights: `Wan-AI/Wan2.1-T2V-1.3B` @ `37ec512624d61f7aa208f7ea8140a131f93afc9a` and
`Efficient-Large-Model/LongLive-1.3B` (`models/longlive_base.pt`, `models/lora.pt`) @ `cda9138d93872ef109d9a74c270184eb3a5acc6e`.
- Config: `configs/longlive_interactive_inference.yaml`, unmodified except `data_path`, `output_folder`,
`num_output_frames` (= 21 × shot count latent frames) and `switch_frame_indices` (21, 42, …). Denoising steps
`[1000, 750, 500, 250]`, 3 latent frames per block, local attention 12, sink 3, global sink on, LoRA rank 256.
- Frames: Wan VAE decodes latent 0 to frame 0 and latent i ≥ 1 to frames 4i−3…4i, so an episode has 84 × N − 3 frames
(333 / 501 / 669) and shot k (from 1) covers frames `[84(k−1) − 3, 84k − 3)` (shot 1 starts at 0, the last ends at the
final frame): 4-shot ranges are `[0,81) [81,165) [165,249) [249,333)`.
- 832×480, 16 fps, written by the unmodified `torchvision.io.write_video(fps=16)`.
- Seed 1 (config default). LongLive seeds once per process; one process ran each shot-count config (4 → 6 → 8 shots)
in priority order.
- Attention backend: FlashAttention-2 (`flash_attn` 2.8.3.post2). torch 2.8.0+cu128. NVIDIA H200: GPU 1 alone, then GPUs 0-3 (claim-based workers) after the ShotStream evaluation finished; `gpu_index` per episode in metadata.json and timing.csv.
- Inference script: `interactive_inference.py` was copied to `interactive_episode.py` with the changes in
`runs/interactive_episode.diff`: output named `<episode_id>.mp4`; skip an episode whose output already exists with
the correct frame count (its noise is still drawn, so a resumed run matches an uninterrupted one); timing and
peak-memory log lines. Model loading, `pipeline.inference` and `write_video` are unchanged.
- The first launch (2026-11-16 03:58 Dubai) was killed by an instance restart before any episode finished
(`runs/logs/run_killed_by_instance_restart_20261116.log`); all 200 episodes come from the relaunch.
## metadata.json fields
`shot_frame_ranges` ([start, end) frames from 0, contiguous), `cut_frames` (frame where each later prompt takes over),
`cut_type`, `prompt_switch_latent_frames`, `latent_frames`, `num_frames`, `fps`, `seed`, `shots[]` (`prompt`,
`prompt_sha256`, `umt5_tokens`), `video_sha256`, `generation_seconds`, `peak_mem_gb`, `gpu_index`,
`generated_at_dubai`.