FeyNobg 抠图 / FeyNobg Matting

If you've used ComfyUI for any amount of time, you know the background-removal drill: BiRefNet is the default, hair survives, and the best results come from a model that was never built for the job. FeyNobg 抠图 / FeyNobg Matting is that lineage pushed one step further - a node wrapping feyn's open FeyNobg model (Apache-2.0), which takes BiRefNet and adds gradient attention, image-patch injection, and multi-scale input on a Swin-Large backbone. No prompt, no click on the subject, no trimap. Drop an image in, get an alpha out.

The 1.05GB weights auto-download to ComfyUI/models/nobg/FeyNobg on first run, so install is effectively zero-work if you have the pack.

What it's for, and what it isn't

FeyNobg is the "find the main subject and strip the background clean" node. Fast, one step, ideal for batch-processing photos or character renders where the subject is obvious. Its sibling in the same pack - Lucida - is the one that deliberately keeps glass, text, and semi-transparent stuff. They're complementary, not competing: the README's comparison shows FeyNobg isolating a character while Lucida keeps the clouds as foreground. Decide by what your material needs.

Inputs that matter

• image - batches fine; it processes per-frame to keep the VRAM peak down. RGBA input silently drops the alpha and uses RGB.
• resolution - default 1024, the model's native training size. Lower saves VRAM but coarsens edges; higher doesn't necessarily help and can break structure.
• precision - fp32 is the default, but the README measured fp16 output as identical (same alpha mean, 0.657) at half the VRAM and visibly faster. On a 12GB card, use fp16.
• device - auto / cpu. CPU is ten-plus times slower; only for genuinely out-of-VRAM situations.

Optional: alpha_threshold (0.5), alpha_softness (1.0 = no-op), keep_aspect_ratio, invert_mask.

Outputs: alpha (MASK) and cutout (IMAGE on a black background). For a transparent PNG, wire alpha into a JoinImageWithAlpha-style node.

The "whole subject went translucent grey" rescue

The model outputs ~0.5 mid-values for areas it's unsure about, which shows up as the subject looking faint and see-through. The model itself has no inference-time dial for this, so the node gives you a post-process level adjustment. Set alpha_threshold to 0.35 and alpha_softness to 0.3 - the README measured that cutting the translucent-pixel share from 1.49% to 0.31% while barely moving the subject mean. Two honest caveats from the author: it only rescues regions the model already half-recognized (alpha near zero stays near zero), and smaller softness eats real semi-transparent detail like hair. Set softness to 0 for a hard binary edge - fine for solid subjects, bad for hair or glass.

Where people get burned

The near-black-alpha failure is the scary one: the pack's FeyNobg weights were exported under transformers 5.x with renamed Swin backbone keys, and a naive load maps only 405 of 958 parameters - the model runs, doesn't error, and outputs almost nothing. This node renames the keys to match your environment and validates strictly, aborting on any mismatch rather than silently degrading. The takeaway is that you should never see all-black alpha here; if you do, it's a transformers-version mismatch, and switching the environment rather than the node is the fix.

For extreme aspect ratios (phone screenshots past 1:2), the fixed 1024×1024 input squishes them - flip keep_aspect_ratio on for letterboxed edge-padding instead.