---
license: cc0-1.0
task_categories:
- text-to-video
- image-to-video
tags:
- animatediff
- lora
- game-assets
- sprite
- character-animation
size_categories:
- 1K<n<10K
---

# SpriteID-5K v2.0 — AnimateDiff LoRA Training Dataset

Pose-annotated 2D game sprite triplets for sprite-domain motion training.
Built for the BitSoul project.

## v2.0 changes from v1.x

- **Action labels from middle target frame**, not reference frame
  (v1's reference labeling caused 55.9% idle dominance)
- **Per-action quotas enforced during collection** — no post-hoc rebalancing needed
- **Stricter JSONL tag filter** — assets must have animation-specific tags
- **Visual difference floor for idle-labeled triplets** prevents static posing sheets
- **Two-stage motion delta gate** (vs last AND vs reference)

## Dataset Statistics

```text
============================================================
  SpriteID-5K v2.0 — STATISTICS
============================================================
  Total triplets    : 1,758
  Base / flip       : 879 / 879
  Sources           : mgane=802  nyu=956

  Action distribution:
    attack     640  (36.4%)  ████████████████████████████████████
    walk       396  (22.5%)  ██████████████████████
    run        238  (13.5%)  █████████████
    slide      156  ( 8.9%)  ████████
    jump       118  ( 6.7%)  ██████
    idle        98  ( 5.6%)  █████
    hurt        88  ( 5.0%)  █████
    die         24  ( 1.4%)  █
============================================================
```

## Pipeline

- DWPose via easy-dwpose + onnxruntime-gpu
- Action labels via Qwen2.5-VL-7B-Instruct (INT8) on middle target frame
- Skeleton gates: 800px density + coherent bounding box aspect ratio
- Motion gate: 400px vs last accepted, 600px vs reference
- Visual gate: idle triplets require visual_diff ≥ 20
- Hard cap: 12 base frames per sheet, even subsampling
- Augmentation: horizontal flip only