The "Copycat" Update & Macro Vision 🕵️‍♂️🔭

[AndyLone22]

MirrorMetrics v0.10.0 introduces a forensic layer to your LoRA evaluation and solves a critical computer vision limitation regarding close-up shots.

🆕 New Feature: The Copycat Detector
Stop guessing if your model is generalizing or just memorizing.

Nearest Neighbor Analysis: The tool now compares every generated image against your entire training dataset in high-dimensional vector space.

Visual Report: Generates a new HTML report (CopycatReport.html) displaying the "Source Inspiration" next to the generated output.

Overfitting Alarm: Quickly spot if your model is "photocopying" specific training images (Similarity > 0.90) instead of learning the concept.

🛠️ Core Engine Update: "Macro Rescue" (Smart Padding)
We fixed a known limitation of retinaface/InsightFace where extreme close-up shots (macro photography) were often ignored due to the face filling the entire frame.

Fallback Mechanism: If a face is not detected in standard mode, the engine now automatically applies a "zoom-out" padding strategy and retries detection.

Data Recovery: Expect to see ~10-15% more valid images recognized in your dataset, specifically those crucial high-texture close-ups needed for accurate Age/Skin evaluation.

⚡ UX Improvements
Progress Bars: Added tqdm visualization for heavy tasks (like Base64 conversion and Vector Search) so you never have to wonder if the script froze.

---

This discussion was created from the release The "Copycat" Update & Macro Vision 🕵️‍♂️🔭.