darask0/krea2-json-bbox-layout-ja-lora

Krea 2 JSON/BBOX Layout + Japanese Manga LoRA

LoRA for Krea 2 that follows a JSON layout (absolute-pixel xyxy boxes for panels, characters, text and manga symbols) and renders Japanese dialogue, speech balloons and manga symbol marks. Trained locally on an RTX PRO 6000 (96GB) from 353 reviewed Japanese manga pages (Gelbooru manga_page / koma tags, general rating), 3000 steps, rank 64.

This public repo contains only the layout LoRA and evaluation sheets. Character LoRAs shown in comparisons are separate local files and are not included. The training dataset is stored in a separate private repo.

Prompt format

{
  "width": 1024, "height": 1024, "bbox_coord_space": "abs", "bbox_format": "xyxy",
  "summary": {"caption": "..."},
  "panels": [{"id": "p1", "reading_order": 0, "bbox": [16, 16, 1008, 640]}],
  "items": [
    {"label": "character", "character_id": "c1", "caption": "...", "bbox": [80, 80, 560, 600], "panel_id": "p1"},
    {"label": "text", "kind": "dialogue", "caption": "Japanese comic lettering", "text": "明日は遠足だよ！", "bbox": [630, 90, 950, 200], "panel_id": "p1"}
  ]
}

Use a long context (sequence_length 5120 in our tests) so long JSON is not truncated. Character-specific LoRAs can be attached per region with "lora": "<alias>" on character items (regional application is an inference-side feature of the companion codebase).

Evaluation (visual, single seed)

Findings: rough position/size direction, Japanese dialogue, balloons, focus/speed lines and emotion marks are reproduced. Exact box containment, sound-effect letter counts and fine character identity are not guaranteed. See the companion evaluation notes in this repo's files.