v1.3 — Lucida v13: background purity + transparency milestone

Update (2026-07-24): the v13 weights described below were published and then reverted to v7 — community testing surfaced a regression our synthetic design test set had missed: on real-world layered artwork (poster/collage-style illustrations) v13 erases or fades design elements that v7 keeps correctly. The code, benchmark tooling and training infrastructure in this release remain current. v14 is in progress with real layered-design training data and a matching benchmark category; the v13 improvements (reduced background haze, the transparency milestone) will return with it.

Lucida v13 (reverted — see note above)

The headline: overall MAE 0.0250 — the lowest of every model we measure (203 images, 9 categories), and transparency past the commercial reference for the first time (0.0338 vs Ideogram 0.0343).

What changed

• Background purity. The gray haze on real-photo backgrounds reported by the community (the cat masks) drove this whole release. The durable fix was a new background-purity hinge loss — a constant-gradient penalty on the eroded true-background region, gated per sample so glow/glass/design keep their soft supervision (training/torch_losses.py). Mean background residue drops from 0.0096 (v7) to 0.0091, with the faint-haze share cut the most.
• Checkpoint soup. Six controlled epochs were trained and benchmarked; the v13 weights were the uniform average of the two best-balanced checkpoints of the lineage.
• Category records: camouflage 0.0227, complex 0.0465, illustration 0.0082, thin 0.0321; hair held exactly at v7's level.
• Traded ground: fx 0.0180 → 0.0211, text 0.0091 → 0.0103, design 0.0235 → 0.0254 — and, as it turned out, more than the numbers suggested on real-world layered artwork (the regression that triggered the revert).
• New benchmark metrics: bg_mae / bg_smear measure residue over the eroded true-background region. The failed experiments (an alpha-squared compositing bug among them) are documented in the README's training story.

Weights

Currently serving v7: https://huggingface.co/egeorcun/lucida