img2rig

Turn a single AI-generated illustration into a breathing, layered character rig — fully agent-driven.

Mira, the bundled demo character: one txt2img illustration → 13 layers → idle / hit / stagger / enrage, all procedural. Try it without installing anything GPU-side: img2rig preview --spec examples/mira/mira.yaml, or interactively in the browser via runtime/web/

One text prompt in; out comes a character that breathes, blinks, sways her hair on spring physics, staggers when hit, and swings full-frame attack keyframes — with zero manual rigging and zero manual mask painting. The pipeline replaces both the rigger and the cleanup artist with an LLM agent running a visual loop over local models (Stable Diffusion + GroundingDINO + SAM), and the runtime is two dependency-free C++ headers.

prompt ──> [1] generate ──> [2] variants ──> [3] split ──> [4] cleanup ──> [5] export ──> [6] runtime
            txt2img            img2img         DINO+SAM      agent loop      layer PNGs     springs,
            candidates         states &        text masks    point tables,   + rig.txt      breath,
            + contact          pose            + PSD         inpaint, QA                    blinks,
            sheets             keyframes                     sheets                         strikes

Why this exists

• Battle-grade 2D character rigging has no production-level automation — mesh-deformation tools are manual by design. But most of what makes a character feel alive on screen (timing, physical lag, impact) doesn't need mesh warping: rigid layers + springs + full-frame pose keyframes cover it, and that pipeline automates end to end.
• The "manual cleanup" in AI layer separation is really two jobs — the clicker (interactive segmentation prompts) and the QA eye. Both are visual judgment loops, and an agent with vision runs them: look at the probe → edit the point table → rerun → look at the overlay. This repo is that loop, made reproducible: all state in one YAML spec, all observations on disk as images. See docs/agent-playbook.md.

Quickstart

Requirements: Python ≥ 3.10, a local AUTOMATIC1111-compatible WebUI with `-