Turn 2+ photos into a high-fidelity Codex companion — locally.
v0.3: Neural Local Studio. No hatch-pet. No OpenAI API key. Open-weight multi-reference editing when your GPU can run it.
The deterministic v0.2 engine was reliable and private, but it could only warp supplied pixels. It could not genuinely invent a new waving hand, a natural running stride, a faithful kitten pose, or a polished 3D chibi redesign.
v0.3 adds a real local generative engine. When supported hardware is present, PocketMen uses an open-weight multi-reference image editor to create a canonical master and state-specific animation key poses, then applies deterministic micro-motion, atlas assembly and QA. The result is far closer to a modern hosted image editor while remaining independent from OpenAI ImageGen.
2+ reference images
↓
Codex visual identity lock
↓
PocketMen hardware doctor
↓
┌─────────────────────────────────────┐
│ Neural Local Studio │
│ default: FLUX.2 [klein] 4B │
│ optional: Qwen-Image-Edit-2511 │
│ multi-reference identity editing │
└─────────────────────────────────────┘
↓
canonical master + semantic key pose per state
↓
flat chroma extraction / alpha cleanup
↓
deterministic micro-animation
↓
8 × 9 transparent Codex atlas
↓
validation + contact sheet + GIF previews
↓
pet.json + spritesheet.webp
↓
Codex local install
If the neural runtime or compatible GPU is unavailable, PocketMen falls back to the deterministic v0.2 engine instead of asking for OPENAI_API_KEY.
- local text-to-image and image editing;
- multi-reference input;
- Apache-2.0 weights;
- suitable for consumer NVIDIA GPUs;
- selected automatically on compatible hardware.
- multi-image editing;
- strong character consistency and viewpoint editing;
- Apache-2.0;
- heavier than the default backend;
- install explicitly when identity fidelity matters more than setup size or latency.
PocketMen intentionally does not make FLUX.2 [dev] the default because its model license is non-commercial.
draft: one neural canonical master, deterministic nine-state animation.balanced: canonical master + state-specific neural key poses; left-running may derive from right-running to reduce drift.max: canonical master + independently generated key pose for every state.
Upload at least two images and ask:
Use $pocketmen-with-you to create and install a companion from these references.
Use Neural Local Studio when my hardware supports it.
Style: hero-chibi.
Subject type: person.
Quality: max.
Preserve identity strictly and do not use hatch-pet or any OpenAI API key.
The skill should visually summarize stable identity traits from the references and pass them through --identity-notes.
Hardware check:
pocketmen doctorCreate with automatic engine selection:
pocketmen create \
--reference photo-1.jpg \
--reference photo-2.jpg \
--name "Mochi" \
--subject-type animal \
--style soft-real \
--engine auto \
--quality balanced \
--output ./pocketmen-output \
--installForce the default neural model:
pocketmen create ... --engine neural --backend flux2-klein-4b --quality maxUse the heavier identity backend:
pocketmen create ... --engine neural --backend qwen-image-edit-2511 --quality maxINSTALL_SKILL.bat installs the skill and creates an isolated runtime. On Windows, the installer checks NVIDIA VRAM; a compatible machine automatically receives the neural dependency profile.
To explicitly prepare and cache the default local neural model:
PREPARE_NEURAL_ENGINE.bat
For Qwen Identity-Max:
PREPARE_IDENTITY_MAX.bat
Model weights are downloaded from their original model hosts into the normal Hugging Face cache; they are not committed to this repository.
PocketMen emits an 8-column × 9-row atlas, 192×208 per cell, 1536×1872 overall. States are:
idle, running-right, running-left, waving, jumping, failed, waiting, running (Codex working, not physical running), and review.
PocketMen narrows the quality gap specifically for reference-driven Codex companion creation by combining strong open-weight image editing, multi-reference identity anchors, structured prompts, dedicated state generation, deterministic alpha extraction and strict atlas QA.
It does not claim bit-for-bit or general-purpose parity with a proprietary frontier model. Broad world knowledge, typography, complex scene reasoning and difficult multi-object composition can still differ. For the narrow “2+ photos → consistent animated companion” workflow, however, v0.3 is designed to use the strongest local path available before falling back to simple warps.
Raw personal references remain local by default and are never copied into public Git commits/releases. The final install package contains only pet.json and spritesheet.webp.
For operation cadence (24h follow-ups, community sync templates, and weekly reports), use:
docs/maintenance/promo-cycle-playbook.md
PocketMen code: MIT.
Optional model weights retain their own licenses. The default FLUX.2 [klein] 4B and Qwen-Image-Edit-2511 backends are referenced, not redistributed.
PocketMen with You is not affiliated with Nintendo, The Pokémon Company, Game Freak, OpenAI, Black Forest Labs, Alibaba/Qwen, Hugging Face, or GitHub.