quarterturn/nichijou-hd-json-captioned

Nichijou HD Captioned Dataset

1,342 HD anime screencaps from Nichijou with detailed captions, character descriptions, camera angles, and lighting/mood metadata.

Structure

train/
  metadata.jsonl        <- flat file_name + caption pairs
  image_00001.png
  image_00002.png
  ...
captions/
  image_00001.json      <- full detailed metadata per image
  image_00002.json
  ...

The metadata.jsonl has simple flat columns (file_name, caption) for dataset viewers. The individual .json files in captions/ have rich structured metadata per image.

Converting JSON captions to TXT (for LoRA training)

Many LoRA training tools (like kohya_ss) expect .txt files alongside images:

import os
import json

captions_dir = "./captions"
train_dir = "./train"

# Convert JSON captions to TXT and place next to images
for fname in sorted(os.listdir(captions_dir)):
    if not fname.endswith(".json"):
        continue

    json_path = os.path.join(captions_dir, fname)
    txt_name = fname.replace(".json", ".txt")
    txt_path = os.path.join(train_dir, txt_name)

    with open(json_path) as f:
        data = json.load(f)

    # Use subject_and_action as the main caption
    caption = data.get("subject_and_action", "")
    if caption:
        with open(txt_path, "w") as f:
            f.write(caption)

    print(f"Created {txt_name}")

Metadata in each .json file (in captions/)

Field	Description
file_name	Image filename
rating	Content rating (sfw)
medium	Medium type (anime_screencap)
subject_and_action	Detailed scene description (use as caption)
style_description	Object: mood, lighting, time_of_day, camera, color_palette
characters	Array of character objects (name, description, pose, etc.)
captured_text	Array of text detected in the image

Each characters object contains: name, source, description, state_of_dress, visible_body_parts, action, gaze, lighting_on_subject.