FL Depth Blur

Most "blur the background" nodes in ComfyUI just gaussian-blur based on a flat mask - one focal plane, one blur amount, done. FL Depth Blur is the version that actually reads distance: hand it your image plus a depth map from any depth estimator upstream in your graph, and it produces a proper lens-like falloff - sharp at your chosen focal depth, progressively blurrier the further a pixel sits from it in either direction.

The mechanism

It precomputes a set of blur kernels (gaussian, box, or disc - disc is the one that gives you that hexagonal bokeh-shaped out-of-focus look real lenses produce), builds a per-pixel "how blurry should this be" map from your focal depth and a falloff curve, then lerps between the sharp and blurred layers using that map. It's the same conceptual approach a compositor uses for defocus, just running on a diffusion-pipeline depth map instead of a 3D renderer's Z-buffer.

Inputs and outputs that matter

image and depth_map are both required IMAGE inputs - the depth map is doing the actual work, so a bad depth map means a bad blur no matter what else you tune. focal_depth (0-1, default 0.5) sets where in the depth range stays sharp; focal_range (0-1, default 0.1) is how wide that sharp zone is before falloff kicks in; max_blur (0-64, default 12) caps how blurry the farthest pixels get. falloff_curve (linear / quadratic / smooth / exponential) shapes how aggressively blur ramps with distance - quadratic, the default, reads more natural than linear. blur_mode picks gaussian/box/disc as above.

near_blur_strength and far_blur_strength (both 0-1) blur the foreground and background independently - useful for a sharp subject with a soft background where you don't want stuff drifting toward the lens to blur symmetrically. depth_invert flips which end of the map counts as "close," which matters because different depth estimators disagree on convention. depth_remap_low/depth_remap_high rescale the usable depth range if your map is compressed into a narrow band. bokeh_brightness_boost (1-4) punches up bright spots in the blur to fake blown-out lens highlights. mask_feather softens the edge of the optional protect_mask input if you want to hard-protect a region - a face, say - from blurring regardless of its depth value.

Outputs: image (the result), blur_amount_viz (an IMAGE visualizing the blur-strength map - genuinely useful for debugging a region that isn't blurring the way you expect), and in_focus_mask (MASK, marking what landed in the sharp zone).

Installing it

ComfyUI Manager: search Fill-Nodes, install, restart. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/filliptm/ComfyUI_Fill-Nodes, then restart. No model download needed - it's pure image processing on whatever depth map you feed it.

Where people get burned

Depth map quality is the whole game. A noisy or low-contrast depth map produces blotchy, inconsistent blur that no amount of falloff-curve tweaking fixes - check blur_amount_viz before touching any other slider if a result looks wrong. And watch depth_invert early: if cranking focal_depth toward 1.0 blurs your foreground instead of your background, your depth map's convention is flipped relative to what this node expects, and no other slider will fix that either.