Skip to content

💃 Free CC0 mannequin pose reference library for AI image tools (Nano Banana, ChatGPT, ComfyUI, Midjourney) — 250+ poses × 5 angles in line art, names in ja/zh/en/ko

21.0 KB
Raw Blame
"""Shared definitions for the planned scene, prop, furniture, light and animal libraries (docs/<lib>/PLAN.md).

Each library follows the costume method (docs/costume/PLAN.md): category rules in data/<lib>-src/_categories.json,
one <CAT>.json of forms × variants, expanded by series_cards.py into data/<lib>s.json and turned into paste-ready
Muse briefs by series_brief.py. What differs per library lives here: ID letter, views and prompt templates.
"""

LANGS = ("zh", "ja", "en", "ko")

# Backgrounds shared with the costume library: generation only asks for a flat colour, intake re-composites it.
BG = {"light": "uniform flat light grey background with no gradient, no vignette and no floor shadow",
      "dark": "uniform flat mid grey background with no gradient, no vignette and no floor shadow"}

# ---- scenes (E###): an empty location seen from the pose library's three camera heights ----------------------
# Poses are drawn front / diag / side at eye level, above at about 30° down and below at about 45° up
# (docs/GENERATION.md), so one scene needs only those three heights; the camera never turns away from the spot.
SCENE_VIEWS = ("eye", "above", "below")
SCENE_SPOT = ("a clear, unobstructed patch of ground in the centre of the frame about 4 metres from the camera, "
              "big enough for one standing adult, with nothing standing, lying or hanging there")
SCENE_VIEW_TEXT = {
    "eye": "eye-level view: the camera 1.6 metres above the ground, perfectly level, the horizon or the far "
           "wall's eye line exactly at the vertical centre of the image, vertical lines vertical",
    "above": "a high-angle view: the camera raised to 3.5 metres above the ground and tilted 30 degrees downward, "
             "still facing the same direction, the empty standing spot now at the centre of the image and more of "
             "the ground visible",
    "below": "a low-angle view: the camera 0.4 metres above the ground and tilted 35 degrees upward, still facing "
             "the same direction, the ground at the standing spot visible only along the bottom edge and the rest "
             "of the frame showing what rises behind it (walls and ceiling, or trees and sky)",
}
SCENE_FRONT = (
    "Photograph of {desc}. An empty location with nobody in it, {view}, 35 mm lens, deep focus so the whole scene "
    "is sharp, {spot}. {light}. Realistic materials and scale, natural colour, 3:4 vertical, location-scouting "
    "reference photo.\n{rules}\n"
    "No people, no figures, no silhouettes, no mannequins, no statues, no animals, no reflections of people, no "
    "readable text, no signs with lettering, no logos, no brand names, no licence plates, no real landmark, no "
    "watermark. {negatives}")
SCENE_TURN = (
    "Image 1 is the eye-level view of a location. Show exactly the same location, facing the same direction from "
    "the same spot, as {view}. Keep these identical: {keep}; same time of day, light, weather and season, same "
    "colours and materials. The empty standing spot stays empty. Add nothing image 1 does not support beyond what "
    "the new camera height reveals (more floor from above, more ceiling or sky from below). 35 mm lens, 3:4 "
    "vertical. No people, no figures, no animals, no readable text, no logos.")

# ---- props (T###): one object, product photograph, same four views as costumes -------------------------------
PROP_VIEWS = ("front", "back", "side", "diag")
PROP_VIEW_TEXT = {
    "back": "a straight back view showing the full back of the object",
    "side": "a true 90-degree side profile showing the object's left side, with its front pointing toward "
            "image-left, showing its depth and thickness",
    "diag": "a three-quarter view with the camera moved 45 degrees around toward the object's left side, so the "
            "front and the left side are both visible and the front turns slightly toward image-left",
}
PROP_REFS = {"back": "Image 1 is the front view.",
             "side": "Image 1 is the front view and image 2 is the back view.",
             "diag": "Image 1 is the front view and image 2 is the side view."}
# float: held or worn objects hang in mid-air in their natural orientation; stand: furniture and other things that
# rest on the ground stand upright on an invisible floor, so the bottom edge is level.
PROP_PLACE = {"float": "floating in mid-air in its natural upright orientation with nothing holding it",
              "stand": "resting in its natural position on an invisible level floor (upright, or lying flat where "
                       "stated), its feet or base on a flat line, with no visible floor"}
PROP_FILL = {"S": "about 50%", "M": "about 65%", "L": "about 80%"}
PROP_FRONT = (
    "Product photograph of {desc}, a single object {place}, straight front view, camera level with the middle of "
    "the object, centered, the object filling {fill} of the frame height, soft even studio lighting, {bg}, "
    "accurate material colour and texture, visible construction details such as joints, seams, hardware and "
    "edges, 3:4 vertical, catalogue style.\n{rules}\n"
    "No person, no hands, no mannequin, no display stand or hook (the object's own built-in stand or tripod stays), "
    "no objects other than those stated, no text, no logo, no brand mark, no watermark. {negatives}")
PROP_TURN = (
    "{refs} They are product photographs of one object. Show exactly the same object from {view}. Keep these "
    "identical: {keep}; same proportions, colours, materials and details. Add nothing the reference images do not "
    "support. {back_rule}Same background, lighting, scale and framing, 3:4 vertical. No person, no hands, no "
    "text, no logo.")
PROP_BACK_RULE = ("The back is the object's real back, not a mirror of the front: front-only details such as "
                  "screens, dials or decorations do not continue onto the back. ")

# ---- furniture (F###): what a posed figure sits, leans or lies on --------------------------------------------
# Split out of the prop library because furniture is composited *under* a figure: it needs the pose library's
# high-angle view (seat and table tops) and real dimensions, so intake can scale every piece to one px/cm and a
# figure sits at the right height. Below is skipped: furniture seen from the floor is rarely useful.
FURN_VIEWS = ("front", "back", "side", "diag", "above")
FURN_VIEW_TEXT = {
    "back": "a straight back view showing the full back of the piece",
    "side": "a true 90-degree side profile showing the piece's left side, with its front pointing toward "
            "image-left, showing seat depth, height and the profile of the back and legs",
    "diag": "a three-quarter view with the camera moved 45 degrees around toward the piece's left side, so the "
            "front and the left side are both visible and the front turns slightly toward image-left",
    "above": "a high-angle view from about 30 degrees above, in front and slightly to the left (the pose library's "
             "above view), so the seat, top surface or mattress is clearly visible",
}
FURN_REFS = {"back": "Image 1 is the front view.",
             "side": "Image 1 is the front view and image 2 is the back view.",
             "diag": "Image 1 is the front view and image 2 is the side view.",
             "above": "Image 1 is the front view and image 2 is the three-quarter view."}
FURN_FRONT = (
    "Product photograph of {desc}, real size about {w} cm wide, {d} cm deep and {h} cm high, a single piece of "
    "furniture standing on an invisible level floor with its feet or base on a flat line and no visible floor, "
    "straight front view, camera level with the middle of the piece, centered, the piece filling about 75% of the "
    "frame width or height, whichever is reached first, soft even studio lighting, {bg}, accurate material colour "
    "and texture, visible joinery, upholstery seams and hardware, 3:4 vertical, catalogue style.\n{rules}\n"
    "No person, no mannequin, no hands, no floor, no rug, no wall, no shadow on a floor, no other furniture, no "
    "objects on or around it unless stated, no text, no logo, no brand mark, no watermark. {negatives}")
FURN_TURN = (
    "{refs} They are product photographs of one piece of furniture. Show exactly the same piece from {view}. Keep "
    "these identical: {keep}; same proportions, real size, colours, materials, number of legs, slats and "
    "cushions. Add nothing the reference images do not support. {back_rule}Same background, lighting and scale, "
    "3:4 vertical. No person, no floor, no text, no logo.")
FURN_BACK_RULE = ("The back is the piece's real back, not a mirror of the front: upholstered backs show their "
                  "back panel, open frames show the back of the joinery. ")

# ---- animals (A###): probe only until the style is decided (docs/animal/PLAN.md, D-14) -----------------------
# Same five camera positions as the pose library, so an animal reference can sit beside a pose reference.
# The first image of each item is its `main` view and every other view references it. Quadrupeds use the side:
# lying, stretching and running hide their key shape (hip height, leg phase) in depth, and both probes drew them
# wrong from the front; birds and the coiled dragon read best from the front.
ANIMAL_VIEWS = ("front", "diag", "side", "above", "below")
ANIMAL_VIEW_TEXT = {
    "front": "straight front view, the camera level with the animal's shoulder and facing its chest",
    "diag": "three-quarter view, the camera moved 45 degrees around toward the animal's left side, level with "
            "the shoulder",
    "side": "true side profile of the animal's left side, the head pointing toward image-left, camera level with "
            "the shoulder",
    "above": "high-angle view from about 30 degrees above, in front and slightly to the left, so the top of the "
             "head and the whole back are visible",
    "below": "low-angle view from about 45 degrees below, in front and slightly to the left, so the underside of "
             "the jaw, chest and belly are visible and the nearest paws look larger than the head",
}
# Style A follows the pose library's line art: one smooth organic contour with light construction lines, so one
# card serves every breed and coat. probe-all01 asked for a "wooden artist's mannequin" with a "segmented neck" and
# got robots (tube segments, bolts, coiled-spring tails), so the prompt now names the drawing style instead of a
# jointed toy and forbids the mechanical parts. Style B is a photograph, only tried in the probe to decide D-14.
ANIMAL_LINE = (
    "Clean black-line figure drawing for a pose reference, in the style of an artist's anatomy construction "
    "sketch: an adult {species} drawn as one smooth, continuous organic outer contour that follows the real "
    "anatomy, muscle masses and proportions of the animal, with a few thin light construction lines inside (a "
    "spine line, cross-contour curves over the ribcage and pelvis) and small light circles marking the shoulder, "
    "elbow, hip, knee and hock joints; a smooth neck that flows into the body and {tail}; no fur, no eyes, no "
    "mouth detail, no markings. Pose: {desc}. {view}. Whole body in frame with even margins, thin even black "
    "lines on pure white, no shading, no colour, no ground, no background, 3:4 vertical.\n{rules}\n"
    "Not a robot, toy or wooden puppet: no mechanical parts, no cylinders or tube segments, no bolts, pins, screws "
    "or hinges, no ribbed or segmented neck or tail, no coiled-spring tail. No people, no hands, no other animals, "
    "no text, no watermark. {negatives}")
ANIMAL_PHOTO = (
    "Realistic photograph of an adult {species} with {coat}. Pose: {desc}. {view}. Whole body in frame with even "
    "margins, soft even studio lighting, uniform flat light grey background with no gradient, no floor shadow, "
    "3:4 vertical.\n{rules}\nNo people, no hands, no collar, no clothing, no props, no other animals, no text, no "
    "watermark. {negatives}")
# Order after the main view: the opposite flat view (front for a side main, side for a front main), then the three
# in-between views. probe-r3-01 turned a lone side reference into a "three-quarter" that stayed a flat profile, so
# the in-between views reference both flat views, as the prop library does.
ANIMAL_ORDER = {"side": ("side", "front", "diag", "above", "below"),
                "front": ("front", "side", "diag", "above", "below")}
ANIMAL_REFS = {"one": "Image 1 is the {main} view of a pose reference.",
               "two": "Image 1 is the {main} view and image 2 is the {second} view of the same pose reference."}
ANIMAL_TURN = (
    "{refs} Draw exactly the same animal in exactly the same pose from {view}. Keep these identical: {keep}; the "
    "same style, line weight and proportions. The pose must not change, only the camera moves; do not mirror the "
    "references. No people, no text.")

# ---- light presets (L###): one light setup on a fixed neutral subject, three eye-level camera positions ---------
# Light is an attribute, not an object: the subject never changes within a category, so the only variable is the
# light. The light stays fixed in the world while the camera moves, which shows how it wraps the subject; each
# preset's `desc` is also the sentence scenes and the combinator reuse.
# Two subjects. `figure` (the pose library's faceless mannequin, full length) reads light on a body in a space. It
# cannot read portrait patterns: at full length the head is tiny and an egg-shaped head has no nose to cast the
# loop or Rembrandt shadow, so the studio positions (PO) use `bust`, a plaster cast with facial planes but no eyes,
# mouth or hair, framed head and shoulders. Both carry a grey sphere, the photographer's standard light reader.
# probe-r2-01: naming a female chest (even "no nipples") got three indoor presets refused, so the figure is
# described only as a slim fashion-drawing mannequin with a simplified torso.
LIGHT_VIEWS = ("front", "diag", "side")
LIGHT_SUBJECTS = {
    "figure": {
        "long": "a life-size faceless fashion-drawing mannequin: one smooth seamless matte light-grey shell "
                "with slim, tall adult proportions and a smooth simplified torso, a smooth egg-shaped head with no "
                "features, small "
                "spherical joints only at the shoulders, elbows, wrists, hips, knees and ankles, and hands with "
                "separated fingers, not a wooden artist's dummy: no wood grain, no blocky segments, no hinges, no "
                "screws; standing upright facing the camera with arms relaxed at its sides, "
                "and a matte mid-grey sphere 25 cm across on a slim grey pedestal at hip height, placed one metre in "
                "front of it and one metre to its right (image-left), so the sphere stays visible from the front, "
                "three-quarter and side views",
        "short": "a grey mannequin and a grey sphere",
        "frame": "Whole figure and sphere in frame",
        "read": "on the mannequin and the sphere: which side is lit, where the highlight sits on the sphere, how "
                "soft the shadow edge is and which way the cast shadow falls",
        "views": {
            "front": "straight front view, camera at chest height 4 metres away, 50 mm lens",
            "diag": "the camera moved 45 degrees around toward the mannequin's left side (it now stands "
                    "three-quarter to the camera), same height and distance",
            "side": "the camera moved 90 degrees around to the mannequin's left side, showing it in profile facing "
                    "image-left, same height and distance",
        },
        "negatives": "No people, no real face, no hair, no clothing, no wooden artist's dummy, no blocky or hinged joints, no text, no logo, no watermark.",
    },
    "bust": {
        "long": "a life-size matte light-grey plaster head-and-shoulders drawing cast of a generic adult on a grey "
                "plinth, facing the camera, with simplified sculpted planes only: forehead, brow ridge, smooth "
                "shallow eye sockets, a straight nose bridge and tip, cheekbone planes, a smooth closed area where "
                "the mouth would be, jaw, chin, neck and shoulders, a smooth bald skull, uniform matte plaster, and "
                "a matte mid-grey sphere 15 cm across on a slim grey stand at shoulder height, 40 cm in front of "
                "the cast and 40 cm to its right (image-left), so the sphere stays visible from the front, "
                "three-quarter and side views",
        "short": "a grey plaster head-and-shoulders cast and a grey sphere",
        "frame": "Head, shoulders and the sphere in frame, the head filling the upper half",
        "read": "on the face planes and the sphere: which side of the face is lit, the shape and direction of the "
                "nose shadow, any lit patch on the shadow-side cheek, how soft the shadow edge is and where the "
                "highlight sits on the sphere",
        "views": {
            "front": "straight front view, camera at the height of the cast's eye sockets 1.8 metres away, 85 mm "
                     "lens",
            "diag": "the camera moved 45 degrees around toward the cast's left side (it now faces three-quarter "
                    "to the camera), same height and distance",
            "side": "the camera moved 90 degrees around to the cast's left side, showing the head in profile "
                    "facing image-left, same height and distance",
        },
        "negatives": "No people, no real or recognisable face, no eyes, no eyebrows, no lips or mouth opening, no "
                     "hair, no skin texture or make-up, no expression, no clothing, no text, no logo, no "
                     "watermark.",
    },
}
LIGHT_FRONT = (
    "Lighting reference photograph: {subject}, {setting}. Lighting: {desc}. {view}. {frame}, deep focus, "
    "true-to-life exposure with no added grading, 3:4 vertical. The light must read clearly {read}.\n{rules}\n"
    "{subject_negatives} {negatives}")
LIGHT_TURN = (
    "Image 1 is a lighting reference photograph of {short}. Show exactly the same moment from a new camera "
    "position: {view}. The lights, the subject, the sphere and the surroundings do not move; only the camera moves, "
    "so the lit side, highlight and cast shadow now appear where that light would fall from the new angle. Keep "
    "these identical: {keep}. Same exposure and colour, 3:4 vertical. No people, no text. Do not mirror image 1.")


def light_front(subject: str) -> str:
    """The front template with one subject filled in; {setting} {desc} {rules} {negatives} stay per item."""
    s = LIGHT_SUBJECTS[subject]
    return (LIGHT_FRONT.replace("{subject}", s["long"]).replace("{view}", s["views"]["front"])
            .replace("{frame}", s["frame"]).replace("{read}", s["read"])
            .replace("{subject_negatives}", s["negatives"]))


def light_turn(subject: str) -> str:
    return LIGHT_TURN.replace("{short}", LIGHT_SUBJECTS[subject]["short"])


LIBS = {
    "scene": {"letter": "E", "src": "data/scene-src", "out": "data/scenes.json", "views": SCENE_VIEWS,
              "label": "场景库", "plan": "docs/scene/PLAN.md", "matte": False,
              "statuses": ("content", "spot", "view", "consistency", "form")},
    "prop": {"letter": "T", "src": "data/prop-src", "out": "data/props.json", "views": PROP_VIEWS,
             "label": "道具库", "plan": "docs/prop/PLAN.md", "matte": True,
             "statuses": ("content", "matte", "view", "consistency", "form")},
    "furniture": {"letter": "F", "src": "data/furniture-src", "out": "data/furniture.json", "views": FURN_VIEWS,
                  "label": "家具库", "plan": "docs/furniture/PLAN.md", "matte": True,
                  "statuses": ("content", "matte", "view", "consistency", "form", "scale")},
    "light": {"letter": "L", "src": "data/light-src", "out": "data/lights.json", "views": LIGHT_VIEWS,
              "label": "光线库", "plan": "docs/light/PLAN.md", "matte": False,
              "statuses": ("content", "subject", "light", "turn")},
    "animal": {"letter": "A", "src": "data/animal-src", "out": "data/animals.json", "views": ANIMAL_VIEWS,
               "label": "动物库", "plan": "docs/animal/PLAN.md", "matte": False,
               "statuses": ("content", "style", "limbs", "view", "consistency", "form")},
}

# Image tiers once a view passes QC (series_build.py, verify.py). Matted libraries (prop, furniture) follow the costume
# library: house grey #D9D9D9 at q60 in three widths (D-10 7); the transparent cut-out stays in the source set (D-18).
# The others keep their own background and use the pose tiers. Settings checked side by side on run01 samples,
# 2026-11-01.
LOOKS = {
    True: {"grey": {"w300": (300, 400), "w600": (600, 800), "w900": (900, 1200)}},
    False: {"plain": {"w300": (300, 400), "w600": (600, 800), "w900": (900, 1200)}},
}
CWEBP = {"grey": ["-q", "60", "-m", "6", "-sharp_yuv"],
         "plain": {"w300": ["-q", "70", "-m", "6"], "w600": ["-q", "72", "-m", "6"], "w900": ["-q", "75", "-m", "6"]}}
STATUSES_ALL = ("pass", "missing")


def qc_paths(lib: str) -> tuple:
    """(qc input dir, qc json) for a library, e.g. data/prop-qc-input, data/prop-qc.json."""
    return f"data/{lib}-qc-input", f"data/{lib}-qc.json"