Pocketmen With You

PocketMen with You — Neural Local Studio

Create a Codex companion from at least two user-provided reference images. The normal workflow is self-contained and must not invoke $hatch-pet or request OPENAI_API_KEY.

Product goal

The user should be able to upload 2+ images and get a polished, identity-consistent Codex companion with minimal setup. PocketMen should aim for hosted-image-editor-like quality within this narrow companion-creation task, while being honest that local open-weight models are not guaranteed to equal a proprietary frontier model on every visual task.

Engine order

Use this decision order:

1. inspect the references visually and build an Identity Lock;
2. run PocketMen hardware doctor;
3. if compatible neural runtime/hardware is available, use neural-local;
4. if the neural runtime is missing but an NVIDIA GPU with roughly 13 GB+ VRAM is present, bootstrap the skill-local neural profile once, then retry;
5. if neural inference is unavailable or fails, fall back to the deterministic local renderer;
6. never route failure into hatch-pet or OpenAI API-key setup.

Default neural backend

flux2-klein-4b

• open-weight local generation/editing;
• multi-reference editing;
• commercial-friendly Apache-2.0 weights;
• intended for consumer NVIDIA GPUs;
• preferred for most users.

Identity-Max backend

qwen-image-edit-2511

Use only when the user prioritizes identity consistency over setup size/latency or when FLUX state outputs visibly drift. This backend is heavier and should not be silently installed unless needed.

Deterministic fallback

local-deterministic-motion-puppet

Guaranteed no-model path. It preserves supplied pixels and produces a valid Codex atlas, but cannot invent unseen anatomy/poses at neural quality.

First step: visual Identity Lock

Before running any command, inspect all supplied references and summarize only stable visible details.

Record:

• subject_type: person, animal, creature, mascot, or auto;
• face/head shape;
• hair/fur/material and stable markings;
• eye color and eye shape;
• silhouette/body proportions;
• fixed accessories (collar, earrings, glasses, watch, ribbon, backpack, etc.);
• stable outfit details when the user wants them locked;
• personality cues relevant to motion;
• any traits that vary between images and must not be locked.

Turn these into one concise --identity-notes string. This is important: the local neural backend sees image references, but explicit stable identity notes substantially reduce drift across animation states.

Do not infer sensitive identity attributes. Describe only visible appearance needed for rendering.

Style choice

Allowed styles:

• soft-real — photo-faithful humans/animals; preserve fur/hair/eyes/markings/accessories;
• hero-chibi — premium 3D toy-like chibi, handsome/cute balance, roughly 2.7–3 heads tall;
• plush — premium collectible plush interpretation;
• capsule-creature — original PocketMen pocket-creature styling using PocketMen's own red/yellow Companion Capsule;
• auto — conservative; use soft-real unless the user clearly requests stylization.

For a real deceased or memorial pet, prioritize identity fidelity over cuteness and do not imply the digital pet is literally the deceased animal.

Quality choice

• draft: one neural canonical master + deterministic nine-state motion;
• balanced: canonical master + state-specific neural key poses; recommended default;
• max: independently generate every state key pose; best for publication-quality showcases and important personal companions.

Use max when the user explicitly asks for maximum quality, ImageGen-like results, or very high fidelity.

Runtime discovery

Resolve the installed skill directory first.

Runtime Python:

Windows:

<skill-dir>\.venv\Scripts\python.exe

macOS/Linux:

<skill-dir>/.venv/bin/python

If the virtual environment does not exist, run:

python <skill-dir>/scripts/setup_runtime.py --skill-dir <skill-dir> --profile core

Hardware doctor

Run:

<python> <skill-dir>/scripts/create_local_pet.py doctor

If recommended_engine is neural-local but the neural dependencies are missing, run once:

python <skill-dir>/scripts/setup_runtime.py --skill-dir <skill-dir> --profile neural

The first real neural creation may download black-forest-labs/FLUX.2-klein-4B into the normal Hugging Face cache. This is a local model download, not an OpenAI API call, and no OpenAI key is required.

If the user explicitly requests Identity-Max, use:

python <skill-dir>/scripts/setup_runtime.py --skill-dir <skill-dir> --profile identity-max

Creation command — default high-quality path

<python> <skill-dir>/scripts/create_local_pet.py create \
  --reference /absolute/path/ref1.jpg \
  --reference /absolute/path/ref2.jpg \
  --name "<display name>" \
  --pet-id "<stable-id>" \
  --subject-type <person|animal|creature|mascot|auto> \
  --identity-notes "<stable visible identity details>" \
  --style <soft-real|hero-chibi|plush|capsule-creature|auto> \
  --engine auto \
  --backend auto \
  --quality balanced \
  --output /absolute/path/to/output \
  --install

For maximum local quality:

--engine neural --backend flux2-klein-4b --quality max

For identity-sensitive heavy mode:

--engine neural --backend qwen-image-edit-2511 --quality max

Do not ask for API credentials if these fail. If neural fallback is allowed, PocketMen will use the deterministic path and report the reason.

Neural rendering design

PocketMen owns the full pipeline:

1. choose a chroma color maximally separated from the references;
2. use 2–3 references as identity anchors;
3. generate a canonical master in the requested style;
4. generate semantically correct state-specific key poses from canonical + raw references;
5. use a strict structured prompt: purpose → subject → identity lock → action → style → composition → background → constraints;
6. render on a perfectly flat opaque chroma background;
7. remove only chroma-like pixels connected to the frame border, protecting same-colored eyes/accessories inside the subject;
8. use deterministic micro-motion between key poses to reduce identity flicker;
9. build, validate, preview and package the 8×9 Codex atlas.