SpriteID-5K: AnimateDiff LoRA Training Dataset

SpriteID-5K v2.0 is a dataset of 2D game sprite triplets annotated with pose information, designed for training motion models in the sprite domain. It was created by author Mohamed-Yossri for the BitSoul project and was last updated on Hugging Face in May 2026. The dataset's version 2.0 includes changes to action labeling and collection quotas to address imbalances present in the prior version.

Use Cases

• Train motion generation models based on pose-annotated sprite triplets.
• Fine-tune animation-specific LoRA adapters using the provided sprite-domain data.
• Benchmark sprite animation synthesis techniques based on the curated action labels.
• Study the impact of dataset balancing on model performance based on the described per-action quotas.

Strengths

• Version 2.0 addresses a specific data imbalance, noting that v1's labeling caused 55.9% idle dominance.
• Per-action quotas were enforced during collection, eliminating the need for post-hoc rebalancing.
• A stricter JSONL tag filter ensures assets have animation-specific tags.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license information are unknown, limiting suitability assessment.
• The dataset's description references a visual difference floor, but the specific threshold is not provided in the available metadata.

Provenance

Source

Hugging Face, uploaded by Mohamed-Yossri.

Collection Method

Collection of 2D game sprite triplets with pose annotations, built for the BitSoul project.

Freshness

Last updated 2026-05-30 16:45:18; freshness should be verified.