pose-blocking
/pose-blocking
Ustala pozę człowieka przed generowaniem lub zdjęciem: kartę pozy, równowagę, stawy, dotknięcia i kamerę oraz szkielety ujęć.
- Autor
- fangdididi ↗
- Repozytorium
- pose-blocking ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 30.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 6
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 30.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Ustala pozę człowieka przed generowaniem lub zdjęciem: kartę pozy, równowagę, stawy, dotknięcia i kamerę oraz szkielety ujęć.
Wyzwalacze z opisu
Oryginalny opis (EN)
Block a human pose like a stage director before you generate or shoot it. Turns a vague pose idea ("leaning on one leg, hand on hip, looking down the street") into a pose card with explicit support, weight, joint angles, touches and camera; checks it for balance, floating supports, impossible joints, missed touches and foreshortening; builds OpenPose-style keypoints and skeleton images in six camera views plus a pose-only prompt; then reviews each generated image against nine named defects and says whether to keep, repaint a region, rerun, or re-block. Use for image/video generation with pose
Jak działa
- What this is
- When to use it
- The pose card
- Steps
- Example
- Tips
Podgląd SKILL.md
---
name: pose-blocking
description: Block a human pose like a stage director before you generate or shoot it. Turns a vague pose idea ("leaning on one leg, hand on hip, looking down the street") into a pose card with explicit support, weight, joint angles, touches and camera; checks it for balance, floating supports, impossible joints, missed touches and foreshortening; builds OpenPose-style keypoints and skeleton images in six camera views plus a pose-only prompt; then reviews each generated image against nine named defects and says whether to keep, repaint a region, rerun, or re-block. Use for image/video generation with pose control, character sheets, illustration and comic reference, photo-shoot posing notes, and animation key poses.
---
# pose-blocking
[中文](SKILL.zh-CN.md)
## What this is
Most bad character images are not bad because of the style. They fail on the body: the weight sits on no foot, a "hand on the hip" lands in mid-air, a left hand becomes a right hand when the figure turns away, a kneeling figure floats a few centimetres above the floor. Text prompts describe a pose loosely, so the model fills in the rest and you burn generations hoping for a lucky one.
Stage directors solve the same problem with *blocking*: before the scene runs, every actor's position, weight and gesture is written down and walked once. This skill does that for a single figure. You turn the idea into a **pose card**, a short JSON file that names what touches the ground, where the weight goes, how each joint bends and where the camera stands. A small standard-library script checks the card like a physical body would. It then gives you three things to steer a generator: a skeleton image, keypoint JSON, and a prompt that describes only the pose. After each generation you fill a nine-line review sheet, and the script tells you the cheapest next move.
What it solves:
- **Physically wrong poses get caught before generation.** Balance, support contact, joint ranges and hand-to-body touches are computed, not guessed.
- **Pose and style stay apart.** The pose block is reused unchanged across retries, so you can see whether a change in style also moved the body.
- **The camera is part of the pose.** The same reach reads well from the side and collapses into a dot from the front. The checker flags limbs that point at the lens, and you can render all six views to pick the readable one.
- **Retries become decisions.** A six-fingered hand needs a local repaint, not a new seed. A figure that keeps drifting from the card needs a simpler card, not more attempts.
## When to use it
Use it when:
- a generated person keeps coming out in the wrong pose, or the same pose is needed across several images or shots;
- you drive a model with pose control (an OpenPose-style ControlNet, a pose or reference input, a depth or 3D-mannequin pass) and need a correct skeleton for it;
- you are writing posing directions for a photo shoot, an illustrator, a comic page or an animation key pose and want them unambiguous;
- the pose has a load-bearing detail such as weight on one leg, sitting at a real chair height, kneeling, a hand on a surface, or a hand touching the face.
Skip it for crowds, tiny background figures, abstract or non-human characters, or when any pose is fine.
### Boundaries
- Work with general poses for illustration, photography, fitness, fashion, animation, comics and games. Do not use this skill to sexualise anyone, and never for content involving minors.
- Do not pose a real, identifiable person in a generated image without that person's consent.
- The body model is a simplified adult figure measured in head heights. It is a planning aid, not motion capture or medical advice. Do not use it to judge whether a stretch or exercise is safe for someone.
## The pose card
One JSON object. Every angle is in degrees and every field is optional except `support`. **`L_` and `R_` always mean the figure's own left and right.**
| Field | Meaning |
| --- | --- |
| `intent` | One line on what the person is doing and why, e.g. "waiting at a bus stop, relaxed". |
| `support` | What carries the body: `L_foot`, `R_foot`, `L_knee`, `R_knee`, `L_hand`, `R_hand`, `seat`, `back`. Add `@height` for a raised surface in head heights, e.g. `seat@1.7` is a normal chair. |
| `weight` | For two-foot stances: `L`, `R` or `even`. |
| `root` | Whole-body tilt: `yaw` (turn, positive turns toward the figure's left), `lean_forward`, `lean_side` (positive toward the left; it also drops the opposite hip). |
| `torso` | Hinge at the hips: `bend_forward`, `bend_side`, `twist` (chest turns toward the left when positive). |
| `head` | `turn` (positive left), `nod` (positive chin down), `tilt` (positive ear toward the left shoulder). |
| `arms.L` / `arms.R` | `raise` (out to the side, 0 hanging, 90 horizontal, 180 overhead), `forward` (swing to the front), `elbow` (0 straight), `elbow_dir` (`front`, `in`, `back`, `up`, `out`: which way the forearm folds). |
| `legs.L` / `legs.R` | `out` (to the side), `forward` (hip flexion; 90 is a seated thigh), `knee` (0 straight; the shin folds backward). |
| `touch` | List of `{"hand": "R_hand", "to": "R_hip"}`. Targets: head, face, chin, neck, chest, belly, L/R_hip, L/R_knee, L/R_hand, L/R_shoulder, L/R_ankle. |
| `hands`, `gaze` | Free text. The script cannot see fingers or eyes, so say what they do. |
| `motion` | `true` when the figure is mid-action (jumping, running, falling). Balance then becomes a note, not an error. |
| `leaning_on` | Name the wall, rail or partner that takes weight outside the base. It has the same effect on balance as `motion`. |
| `camera` | `view`: `front`, `three_quarter`, `side`, `back`, `high`, `low`. `framing`: `full`, `knees`, `waist`, `chest`. |
Root lean tips the whole body, legs included, like a tree falling. Torso bend folds at the hips while the legs stay put. Most "leaning forward" poses are torso bends.
## Steps
1. **Write the intent and the support first.** Ask what touches the floor or a surface and where the weight goes. If you can't answer, you don't have a pose yet. Pick chair and step heights from real furniture: a chair seat is about 1.7 head heights and a bar stool about 3.
2. **Set the big masses, then the limbs.** Start with the pelvis (`root`), then the chest (`torso`), then the head. Limbs come last. A relaxed standing pose almost always has the pelvis tilting toward the free leg and the chest tilting back the other way.
3. **Declare every contact.** A hand on a hip, a chin on a fist or fingers on a shoe goes in `touch`. Any limb that bears weight goes in `support`.
4. **Choose the camera on purpose.** Use the view in which the main gesture is broad and flat to the lens. Reaches toward the camera need the side or three-quarter view, or a depth reference.
5. **Run the checker.** `python3 scripts/pose.py check card.json`. Exit 0 means clean, 1 means notes only, 2 means errors, 4 means the card can't be read. Fix every error by changing one or two angles at a time and run it again. Don't silence a finding by deleting the claim that caused it unless the claim was wrong.
6. **Build the control pack.** `python3 scripts/pose.py build card.json -o out --png`. You get `keypoints_<view>.json` in the common OpenPose 18-point layout, `skeleton_<view>.svg` and `.png` drawn on black, `prompt.txt` and `report.json`. Look at the skeletons for all six views and confirm the hero view reads the way you meant.
7. **Pick a control route.**
- The generator accepts a pose skeleton: give it `skeleton_<hero>.png` (or the keypoints) together with the pose prompt.
- It only accepts a reference image: pose a 3D mannequin or take a phone photo matching the skeleton, then use that.
- It is text only: use the pose prompt, keep the framing generous, and expect more retries. Foreshortened limbs rarely survive text alone.
8. **Keep pose and style in separate blocks.** Paste the pose paragraph unchanged. Identity, clothing, lighting and style go in their own block. Change one block at a time between retries.
9. **Review every keeper with the nine checks** and fill a sheet like `examples/review-sheet.json`:
- `HANDS`: right finger count, no fused or melted fingers, the grip makes sense.
- `FACE`: the face is intact and turned the way the card says.
- `GAZE`: the eyes look where the card says.
- `LIMB_COUNT`: two arms and two legs, nothing extra or missing.
- `MERGE`: no limb passes through the body, a prop or another person.
- `FLOAT`: the supports really touch their surface (contact shadow, compressed cushion, flat sole).
- `CROP`: nothing load-bearing is cut off by the frame in a way that changes the read.
- `DRIFT`: each joint named in the card is roughly where the card puts it.
- `MIRROR`: left and right are not swapped.
Then run `python3 scripts/pose.py review sheet.json`:
- `KEEP` (exit 0): every check passed.
- `LOCAL_FIX` (1): only hands, face or gaze failed, so repaint that region and keep the rest.
- `RERUN` (2): same pose block, new seed, stronger control. Change nothing else.
- `REBLOCK` (3): the pose itself drifted or mirrored again on a second attempt, so simplify the card or pick a clearer view before generating more.
10. **Record what held.** Keep the card, the hero view and the control strength that worked next to the image. That set is your reusable pose for the next shot of the same character.
## Example
The idea was "waiting at a bus stop, weight on one leg, hand on hip, looking down the street".
```json
{
"intent": "waiting at a bus stop, relaxed",
"support": ["L_foot", "R_foot"], "weight": "L",
"root": {"lean_side": -8},
"torso": {"bend_side": 10, "twist": 5},
"head": {"turn": -15, "tilt": -5},
"arms": {"L": {"raise": 8, "elbow": 10},
"R": {"raise": 30, "forward": -10, "elbow": 90, "elbow_dir": "in"}},
"legs": {"L": {"out": 0}, "R": {"out": 14, "forward": 15, "knee": 10}},
"touch": [{"hand": "R_hand", "to": "R_hip"}],
"hands": "right hand loosely on the right hip, left hand relaxed",
"gaze": "down the street to the right",
"camera": {"view": "three_quarter", "framing": "full"}
}
```
A first draft tilted the pelvis the wrong way (`lean_side: 6`), bent the free knee 18 degrees and folded the elbow to 115. The checker answered `SUPPORT_FLOAT: R_foot hangs 0.38 H above its surface` and `TOUCH_MISSED: R_hand is 0.74 H from R_hip`. Tilting the pelvis the other way (`lean_side: -8`, so the right hip drops), nearly straightening the free knee and folding the elbow to 90 fixed both, and `check` now reports `CLEAN`. The built prompt begins: *"Full-body shot, three-quarter view, eye level. Action: waiting at a bus stop, relaxed. Pose: standing with weight on the left leg. Left hip higher than the right, head turned to the right. …"*
The first generation had six fingers on the hip hand and everything else right. The review sheet with `HANDS: fail` returned `LOCAL_FIX`, so only the hand was repainted.
`examples/` also holds a seated reader, a figure kneeling to tie a shoe, and a deliberately broken lean that shows six findings at once.
## Tips
- **Weight first, always.** If the weight claim and the centre of mass disagree, viewers feel the pose is wrong even when they can't say why.
- **One support short is fine on purpose.** Lift a foot out of `support` for a step or a kick, and set `motion` if the body is mid-action.
- **Watch the `FORESHORTENED` note.** A limb pointing at the lens is where text prompts fail most. Change the view or add a depth or mannequin reference.
- **Back views swap sides.** When the figure faces away, write "her own left hand" style wording and check `MIRROR` first.
- **Keep the pose block frozen** while you change style. If the body moves when only the style changed, raise the control strength rather than rewording the pose.
- **Real units help.** One head height is roughly 22 to 24 cm for an adult. Convert furniture to head heights before you set `@height`.
Pliki w paczce (6)
- .gitignore 1 KB
- LICENSE 1 KB
- README.md 6 KB
- README.zh-CN.md 6 KB
- SKILL.md 12 KB
- SKILL.zh-CN.md 12 KB
Wersje
W bazie nie ma innych skilli o nazwie pose-blocking.
Migawka bazy z 13.11.2026 · wersja paczki 676b96403e28 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 86zbrush-pose-presentation
Ustawia pozę modelu i przygotowuje spójne renderowe przeglądy do oceny. - 100avatar-pose-seeds
Zamienia opis pozy w powtarzalne ziarno i mapę rotacji stawów dla awatara humanoidalnego w @three-ws/pose. - 90catimation-director-stage
Steruje scenami 3D w aplikacji CATIMATION: ustawia modele i postacie, animacje, światła, kamery oraz nagrywa ruchy kamery. - 100ai-gallery
Pobiera obrazy referencyjne pozy, głębi, kostiumów i twarzy z galerii do narzędzi generowania obrazów, takich jak GPT Image, Nano Banana czy ComfyUI. - 90reference-role-mapping
Przypisuje referencjom jasne role: temat, pozę, kamerę, otoczenie, paletę i styl, żeby łatwo je zestawiać. - 86concept-art
Generuje koncepty postaci na fal.ai: arkusz modelu z widokiem z przodu i z tyłu oraz pozy na podstawie zapisanej postaci, sprawdzając każdy wynik.