mccoyspace / autocritic

What is autocritic?

autocritic turns multimodal LLMs into art critics by loading JSON "critic cards" distilled from theorists like Wölfflin or Kandinsky—structured summaries of their frameworks for image evaluation. Point it at any PNG via CLI or Python API, and it delivers qualitative feedback (strengths, weaknesses, directives) plus numeric scores on theory-specific axes, like linear vs. painterly. The killer feature: an improvement loop that translates critiques into parameter deltas for generative systems, iteratively refining outputs without rigid mappings—all in Python.

Why is it gaining traction?

Unlike generic "improve this image" prompts, autocritic grounds feedback in actual art theory via ready-to-use cards, yielding consistent, axis-scored evals that reveal autocriticism meaning in visual terms. The CLI shines: `autocritic critique image.png --critic wolfflin` or `autocritic run --iterations 5 --generator rewriter` for looped refinement with contact sheets. LLM-driven param translation adapts to any generator's space dynamically, making critic-card-driven evaluation practical for real workflows.

Who should use this?

Procedural artists steering SVG or L-system generators toward specific aesthetics. AI devs building image pipelines needing structured, theory-based scoring over vague LLMs. Researchers exploring autocritico sinonimo in multimodal setups, like testing Kandinsky's bipolar principles on synthetic art.

Verdict

Worth a spin for Python devs into generative art—CLI and docs make it dead simple despite 17 stars and 1.0% credibility score. Early-stage with strong evals and extensibility via custom cards; fork and iterate confidently under MIT.

(198 words)