akatz-labs/H3-Character-Swap-v2

H3 Character Swap v2 — working collection

Public working collection for collecting source footage and character references before making edited training pairs. Nothing in this initial upload is an approved v2 training example.

Initial inventory: 22 full-length Tears of Steel source videos, 50 synthetic source images, and 75 character references. The videos are 1456×768 at 24 fps, with no audio because the source archives contained images only. There are 20 distinct Tears of Steel shot IDs plus two alternate scaled variants. MOVs use H.264, not ProRes. Full takes are retained; many exceed the intended training window and still need clip selection.

V1 vs V1.5 regression benchmark

Five selected cases show where CharacterSwap Pro V1 loses identity or style and V1.5 preserves it. The pack contains the original paired outputs, four sets of recorded inputs and API graphs, prompts, masks, seeds, checksums, and a replay helper. The frilled-lizard pair includes operator-supplied settings and mask, but its original run is incomplete.

Open the benchmark and replay instructions · Machine-readable cases · Browse all files

Case	Output	Seed	Selected V1 failure
Frilled lizard	960×512, 124 frames / 5.167 s	194729 (operator)	Source actress returns near frame 43
Anime cartographer	960×512, 120 frames / 5 s	194729	Source actor returns near frame 16
Anime archivist	1344×736, 240 frames / 10 s	194731	Identity loss around frames 22–80 and 200–239
Retro orc	1344×736, 120 frames / 5 s	194731	Source face returns after the cut at frame 82
Anime diplomat	1344×736, 120 frames / 5 s	194731	Realistic hair replaces anime styling in the close-up

These are curated qualitative regression cases, not an unbiased evaluation or aggregate performance score. Training overlap has not been audited. Benchmark files are separate from the working collection and are not approved training examples.

Where to add things

• sources/: incoming footage and scene images, grouped by collection. Put new YouTube footage under sources/youtube/<channel>/. Keep original durations here.
• references/: character portraits and character sheets. synthetic-v1/ carries the 70 unique existing references; characters/ holds Akatz, Mantis, the felt creature, Frieren, and Fern.
• samples/: future selected source clips, edited targets, prompts, masks, and review information. Only a template exists so far.
• review/: preview gallery and accept/reject decisions with reviewer notes.
• manifests/: source/reference inventories, checksums, dimensions, frame counts, and provenance. The training sample manifest starts empty.
• licenses/ and LICENSES.md: component attribution and rights information.

Collaborating

1. Choose a stable ID for an incoming asset. Add the file under the appropriate sources/ or references/ folder.
2. Add a matching row to the relevant JSONL manifest, including origin URL, creator, license/permission information, and SHA256. A starter record for future YouTube sources is in sources/youtube/source-template.json.
3. Keep incoming material approved_for_training: false. Collection, clip selection, edit quality, and training approval are separate decisions.
4. Create a sample folder only when you have selected a source interval and references. Copy samples/sample-template.json, retain source IDs and exact frame ranges, and supply the edited target and prompt.
5. Record review notes. Add only approved pairs to manifests/samples.jsonl. Preserve source-group and reference-identity separation when choosing train/evaluation splits; v1 split lineage is retained in the asset manifests.
6. Prefer new filenames/versions to silently replacing someone else's target. Keep previews and rejected experiments clearly marked.

Download the repository:

hf download akatz-labs/H3-Character-Swap-v2 --type dataset --local-dir h3-character-swap-v2

Open review/index.html locally to browse videos and references. This is an asset collection, not yet a trainer-ready Hugging Face dataset schema.

The existing released v1 remains at https://huggingface.co/datasets/akatz-ai/H3-Character-Swap-v1 . This collection does not import its targets or claim new training results.