egeorcun

/

lucida

like 40

Thank you for the feedback. I will check it and get back to you.

Hello, could you please try again. I fixed the issue.

@egeorcun
I think something's wrong, lucida almost always gives smears and leftovers

Thanks for feedback. I'm gonna working on it and i will let you know

@dummy9966 could you please try the same image on v13 for smears and leftovers?

Update: the new weights (v13) are live, and they specifically target the smears you both reported.

What was happening: with a training mix that intentionally contains a lot of semi-transparent ground truth (glass, glow, smoke), the standard matting losses make faint mid-alpha hedging cheap — so the model left a low-alpha haze on real-photo backgrounds that plain MAE barely registers. We added a background-residue metric to our benchmark to make it visible, then trained with a background-purity hinge objective: a constant-gradient penalty over the region the ground truth marks as pure background (eroded, so legitimate soft edges like fur and glass are untouched), gated so the semi-transparent categories keep their soft supervision.

Measured on our 203-image benchmark against ground truth, the released weights vs the previous ones:

• mean background residue: 0.0096 -> 0.0091 overall, with the faint-haze share reduced the most
• overall MAE: 0.0257 -> 0.0250 (still ahead of every model we measure)
• transparency now also passes the commercial reference we benchmark against (0.0338 vs 0.0343)

Being straight about the remaining gap: occasional isolated low-alpha blobs can still show up on busy backgrounds — reduced, not eliminated. That is the top item for the next iteration.

One practical note: if you are testing through ComfyUI's bundled collection, the lucida.safetensors there is an older snapshot — for the current weights please pull from this repo directly (or try the live demo Space). Would appreciate a re-test on your images.

Follow-up: we have temporarily reverted the repo to the v7 weights. Community testing of the v13 build surfaced a regression our benchmark had missed — on layered artwork (poster/collage-style illustrations) it erased or faded design elements that v7 keeps correctly. The background-haze improvements from v13 are real and will return in the next update together with the artwork fix (we are adding real layered-design training data and a matching benchmark category so this class of regression cannot slip through again). Apologies for the churn — reverting felt more honest than leaving a known regression live.