six-nut/PocketMen-with-you

Create high-fidelity Codex companions from 2+ photos with local open-weight neural editing and no OpenAI API key.

What it solves

It allows users to create high-fidelity, animated digital companions (Codex companions) from two or more reference photos entirely on their own local hardware. It solves the limitation of older versions that could only warp existing pixels, enabling the generation of new poses and movements while maintaining character consistency without requiring an OpenAI API key.

How it works

The system uses a "Neural Local Studio" that leverages open-weight image editing models (FLUX.2 [klein] 4B by default, or Qwen-Image-Edit-2511 for higher identity fidelity) to process reference images. It generates a canonical master image and state-specific key poses, then applies deterministic micro-motion and alpha cleanup to assemble an 8x9 transparent sprite atlas. If a compatible GPU is not found, it falls back to a simpler deterministic engine.

Who it’s for

Users who want to create consistent, animated character sprites from personal photos privately and locally, specifically for use as Codex companions.

Highlights

  • Local Generative Engine: Uses open-weight models to create new poses instead of just warping pixels.
  • Multi-Reference Support: Locks visual identity using multiple input photos.
  • Privacy-First: Reference images stay local and are not uploaded to external APIs.
  • Flexible Quality Profiles: Offers draft, balanced, and max settings to control the trade-off between generation speed and pose independence.
  • Automatic Hardware Detection: Includes a "hardware doctor" to select the best available engine based on GPU VRAM.

Related

  • Project
  • Project
  • Project
  • Project