Lucida 抠图 / Lucida Matting

Most background removers quietly fail on the same five things: camouflaged objects, glass, text and logos, VFX glow, and illustration. Lucida 抠图 / Lucida Matting is a model that was trained to win those specific fights. It's a fine-tune of BiRefNet_HR (MIT-licensed, ~885MB, 220M params, Swin-Large backbone), and the author's own 203-image benchmark puts it ahead of a commercial reference on text/logo retention (0.0091 vs 0.0123 MAE) with strong numbers on illustration, camo, and sticker prints.

Like FeyNobg, it's fully automatic - no prompt, no box. First run pulls the weights to ComfyUI/models/lucida/lucida.safetensors automatically.

What it's good at, honestly

This is the node to reach for when "foreground vs background" isn't a clean binary - when keeping the semi-transparent material is the whole point. Glass, smoke, light effects, a logo sitting on a sticker, an illustration with soft edges. The README's side-by-side with FeyNobg makes the philosophy explicit: on a picture of a character with clouds and railings, FeyNobg strips to the character alone while Lucida keeps the translucent clouds as foreground and preserves the railings. Neither is "wrong" - they define foreground differently, and Lucida's definition happens to include transparency. On your own material, run both (the pack even ships a workflow that parallels them) and pick by which cutout matches your intent.

Inputs and outputs

• image - batch-friendly, processed per-frame to cap VRAM. RGBA in drops the alpha channel.
• model_name - weights under models/lucida; auto-downloads if absent.
• precision - default fp16: half the VRAM, and measured nearly identical to fp32. Drop to fp32 only if you get a black frame or artifacts.
• device - auto / cpu (CPU is painfully slow).

Optional: alpha_threshold / alpha_softness (same level-adjust pair as FeyNobg, defaults 0.5/1.0 = no-op), keep_aspect_ratio, invert_mask.

Outputs: alpha (MASK) and cutout (IMAGE on black). Transparent PNG needs a JoinImageWithAlpha-style step downstream.

One thing this node deliberately doesn't have: a resolution selector. The model's decoder is hard-bound to 1024×1024 via patch-split, so unlike FeyNobg there's no 512/768/1280 ladder. You get the native size, and that's that.

The design choice worth knowing about

Most BiRefNet nodes load with trust_remote_code, which means at runtime they pull and execute Python from HuggingFace. This node instead embeds the model code (2,250 lines) inside the pack - pinned version, works offline, auditable - and skips the ImageNet backbone download by constructing with bb_pretrained=False. Weight loading is strictly validated too: any mismatch aborts rather than silently degrading. It's the same careful engineering philosophy as the pack's FeyNobg node, and it's the reason these two are more trustworthy than the average trust_remote_code wrapper.

Troubleshooting

Black or garbage output → switch precision to fp32 and re-run; the fp16 path is solid but a bad environment can trip it. And resist the urge to crush alpha_softness toward 0 on glass or glow material - Lucida's whole reason to exist is preserving real semi-transparency, and a hard threshold throws that away.