CogCanvas — Data Card

A single bundle (CogCanvas.zip) containing the reference images and per-scene prompt/annotation files used to build composite "person + object + background" generation prompts.

---

Top-level layout

CogCanvas_dataset/
├── background/
├── celeb/
├── objects/
└── prompt/

Every file path that appears inside the JSON annotations is relative to the zip root (e.g. celeb/CelebA-HQ-img/13288.jpg, objects/images/airpods_pro_2/img_20.jpg, background/Moc_Chau.png).

---

background/

Flat directory of background scene images.

background/<scene_name>.png

Examples: Ba_Na_Hill.png, Moc_Chau.png, One_Pillar_Pagoda_Hanoi.png. The <scene_name> is what the metadata.json files reference under background_path.

---

celeb/

celeb/
├── CelebA-HQ-img/<image_id>.jpg
├── CelebA-HQ-200_id.csv
└── celeb_metadata.json

• CelebA-HQ-img/<image_id>.jpg — reference portraits. Referenced from prompt annotations as celeb/CelebA-HQ-img/<image_id>.jpg.

• CelebA-HQ-200_id.csv — maps every CelebA-HQ image to its anonymized identity (identity_name such as id_0042).

• celeb_metadata.json — per-identity attributes extracted with QwenVL. Keyed by anonymized id (id_0000, id_0001, …).

  {
    "total_persons": <int>,
    "model": "Qwen/Qwen2.5-VL-7B-Instruct",
    "backend": "transformers" | "vllm",
    "persons": {
      "id_0000": {
        "name":      "id_0000",
        "ref_image": "CelebA-HQ-img/<image_id>.jpg",
        "gender":    "male" | "female" | "unknown",
        "age_group": "child" | "teen" | "young" | "middle-aged" | "elder" | "unknown",
        "body_type": "slim" | "athletic" | "muscular" | "petite" | "average" | "stocky" | "heavy" | "unknown",
        "ethnicity": "black" | "white" | "asian" | "hispanic" | "middle-eastern" | "south-asian" | "mixed" | "unknown"
      }
    }
  }
  

---

objects/

objects/
├── images/<object_dir>/img_*.jpg|png|webp
└── object_metadata.json

• images/<object_dir>/img_*.{jpg,png,webp} — multiple reference shots per object/fashion item. <object_dir> is the slugified product name (e.g. airpods_pro_2, áo_dài_cách_tân_họa_tiết_sen, túi_xách_dior_saddle_bag_blue_oblique_canvas). Referenced from prompt annotations as objects/images/<object_dir>/img_<n>.<ext>.

• object_metadata.json — per-object attributes extracted with QwenVL, keyed by <object_dir>.

  {
    "total_objects": <int>,
    "objects": {
      "<object_dir>": {
        "name":          "<object_dir>",
        "ref_image":     "images/<object_dir>/img_<n>.<ext>",
        "category":      "apparel" | "footwear" | "bag" | "accessory" | "jewelry" | "watch" | "electronics" | "beauty" | "vehicle" | "toy" | "home_appliance" | "sports" | "book" | "art" | "other" | "unknown",
        "primary_color": "red" | "blue" | "green" | "yellow" | "black" | "white" | "gray" | "brown" | "pink" | "purple" | "orange" | "beige" | "gold" | "silver" | "multi" | "unknown",
        "style":         "casual" | "formal" | "sporty" | "traditional" | "luxury" | "vintage" | "modern" | "minimalist" | "streetwear" | "fantasy" | "other" | "unknown",
        "material":      "cotton" | "leather" | "denim" | "wool" | "polyester" | "synthetic" | "metal" | "plastic" | "ceramic" | "wood" | "glass" | "fabric" | "mixed" | "other" | "unknown",
        "gender_target": "male" | "female" | "unisex" | "none" | "unknown"
      }
    }
  }
  

---

prompt/

Per-scene annotations grouped by person count and background, then by generation timestamp:

prompt/
├── total_metadata.json
└── <N_people>/<background_name>/<image_id>/
    ├── dpg.json
    ├── metadata.json
    └── scene_analysis.json

• <N_people> ∈ {two_people, three_people, four_people, five_people} — number of people in the scene.
• <background_name> — matches a file under background/ (e.g. Moc_Chau → background/Moc_Chau.png).
• <image_id> — generation timestamp (YYYYMMDD_HHMMSS).

prompt/total_metadata.json

JSON array — one entry per scene, identical schema to the per-scene metadata.json below, plus a top-level name field. Useful as a flat index over all scenes.

prompt/<N_people>/<background>/<image_id>/metadata.json

Scene-level metadata that ties the prompt to its reference images.

{
  "prompt":            "<final natural-language prompt>",
  "prompt_base":       "<structured / templated form>",
  "prompt_original":   "<pre-anonymization version>",
  "background":        "<background_name>",
  "background_path":   "background/<background_name>.png",
  "no_person":         "two_people" | "three_people" | "four_people" | "five_people",
  "image_id":          "YYYYMMDD_HHMMSS",
  "objects_short":     ["<object name>", ...],
  "objects_person":    ["Person 1", "Person 2", ...],
  "ref_images": {
    "Person 1":        "celeb/CelebA-HQ-img/<image_id>.jpg",
    "<object name>":   "objects/images/<object_dir>/img_<n>.<ext>"
  },
  "person_celeb_mapping": {
    "Person 1":        "id_0099"
  }
}

ref_images keys mix person labels ("Person 1") and object names (matching objects_short). Resolve a person id → attributes via person_celeb_mapping → celeb/celeb_metadata.json.

prompt/<N_people>/<background>/<image_id>/scene_analysis.json

Per-scene structured layout describing each person and their interactions.

{
  "objects": [
    {
      "id":                  "A",
      "person_label":        "Person 1",
      "original_name":       "<descriptor, e.g. teenage girl>",
      "celeb_name":          "id_0099",
      "celeb_gender":        "female",
      "celeb_age_group":     "young",
      "celeb_body_type":     "average",
      "outfit":              "<comma-separated outfit description>",
      "position": {
        "absolute":          "left" | "center" | "right" | "...",
        "relative_to": [
          { "object_id": "B", "relation": "left_of" | "right_of" | "..." }
        ]
      },
      "holding_object":      "<object name>",
      "holding_object_dir":  "<object_dir under objects/images/>"
    }
  ],
  "interactions": [
    /* pairwise interactions between objects[].id values */
  ]
}

object_id cross-references the id field within objects[] of the same file (NOT files under objects/). holding_object_dir resolves under objects/images/<holding_object_dir>/.

prompt/<N_people>/<background>/<image_id>/dpg.json

DPG-style evaluation questions for the scene.

{
  "prompt":            "<scene prompt>",
  "qid2tuple": {
    "1":               "<subject> | <relation> | <target>"
  },
  "qid2dependency":    { "1": ["0"] },
  "qid2question":      { "1": "Is there a person 1?" }
}

Question ids are local to each dpg.json — qid2dependency references those same ids ("0" denotes the root).

---

Path-resolution cheat sheet

Field in JSON	Resolves to (relative to zip root)
background_path	background/<scene_name>.png
ref_images["Person N"]	celeb/CelebA-HQ-img/<image_id>.jpg
ref_images["<object name>"]	objects/images/<object_dir>/img_<n>.<ext>
person_celeb_mapping["Person N"]	key into celeb/celeb_metadata.json → persons
scene_analysis.objects[].holding_object_dir	dir under objects/images/