Skip to content

Depth Lens Blur

Fake a wide-aperture lens after the shot. The subject stays sharp and everything nearer or farther melts into smooth bokeh, according to a depth map.

Depth Lens Blur — before / after
In ComfyUI Add Node → ComfyUI-JH-PixelPro → depth → Depth: Lens Blur
Good for Background blur on portraits, product shots with creamy backgrounds, rescuing busy backgrounds
Speed 2048 × 2048 image arriving from ComfyUI (CPU tensor), depth model excluded: about 0.03 s with a CUDA GPU (RTX 5090, transfers included); about 2 s on CPU only at max_blur 24, up to about 10 s at max_blur 96 with 16 layers

When to use it

On a real lens, only one distance is sharp. The farther something is from that distance, the bigger its blur disc (the circle of confusion). This node does the same per pixel, using a depth map where 1 is nearest.

  • Focus (focus_depth) is the depth that stays sharp.
  • Depth of field (depth_of_field) is the width of the sharp zone. Small values act like f/1.4, large values like f/16.
  • Max blur (max_blur) is how big the blur gets at the farthest distance from focus, in pixels.

No depth map connected? The node estimates one itself with Depth Anything V2 (small model, 518 px). That needs onnxruntime (pip install onnxruntime-gpu).

Quick start

  1. Load Image → Depth: Lens Blur → Preview Image.
  2. Leave depth empty for automatic depth, or connect your own map (for example from Depth Anything V2).
  3. Try depth_of_field 0.1 and max_blur 24, then raise max_blur for stronger blur.
  4. To keep a person sharp automatically, connect a face mask to focus_mask (see below).

Settings

Setting Default What it does Tip
focus_depth 0.5 Depth value that stays sharp (1 = nearest). Ignored when focus_mask is connected.
depth_of_field 0.1 Width of the sharp zone. 0 to 1. Lower = shallower focus.
max_blur 24 Blur radius in pixels at the farthest distance from focus. 1 to 96. Scale up for large images.
bokeh disc disc gives flat, lens-like bokeh. gaussian is a softer blur. disc for photographic look.
layers 8 Number of blur steps blended together. 2 to 16. More = smoother transitions.
blur_foreground on Off: only things behind the focus plane blur. Turn off to keep a near subject edge or foreground sharp.
depth (optional) none Depth map, 1 = nearest. If missing, estimated internally. A map of another size is resized to the image.
focus_mask (optional) none Subject mask. The median depth inside it becomes the focus. Works with face masks or SAM masks.
keep_model_loaded (optional) on Keep the built-in Depth Anything model in VRAM between runs (only used when depth is not connected). Turn off to free its VRAM right after the run; ComfyUI's Free memory button does not reach this model.

Outputs: image (IMAGE) and coc (MASK, blur amount from 0 to 1; multiply by max_blur for pixels), useful for checking what gets blurred.

Chaining: auto-focus on a face

Face Detect Lite face_mask → Depth Lens Blur focus_mask. The face depth is measured automatically and the face stays sharp, however far the person stands. Any other mask works the same way, for example from SAM Mask.

Tips from retouching

  • For video, the automatic depth is normalised on every frame, so the focus can flicker between frames. Feed a steady depth map, or a fixed focus_depth, for clips.
  • Check the depth map first. Blur follows the depth edges, so a wrong depth edge shows as a wrong blur edge.
  • Use max_blur in proportion to the image size: 24 is moderate on a 2K image, 60 or more suits 6K.
  • A little depth_of_field (0.05 to 0.15) keeps the whole face sharp, not just the eyes.
  • Add subtle noise or grain after the blur so blurred areas match the sharp ones.
Under the hood

Circle of confusion per pixel: clamp((|depth − focus| − dof/2) / (1 − dof/2), 0, 1) × max_blur. The node renders layers blur images with radii from 0 to max_blur and blends each pixel between the two nearest radii. Each layer is a normalised convolution (blur(image·w) / blur(w)) where w shrinks for pixels that are sharper than the layer, so a sharp foreground does not smear into the blurred background. Radii above 8 px are blurred on a downscaled copy and upsampled, so memory and time stay bounded (no unfold). The work runs on the GPU when one is available (ComfyUI passes images on the CPU; the result goes back to the same device). Peak GPU memory was about 0.45 GB at 2048 × 2048. The blur map is cleaned up with a small closing (window about max_blur / 8 px) so no thin sharp line appears where the depth crosses the focus plane between two blurred zones; wide in-focus areas stay untouched. NaN values in the depth or focus mask are treated as finite numbers, so one bad pixel cannot spoil the image. Limits: a blurred foreground does not spread over an in-focus background, and depth errors show up as blur errors. Source: core/depth_lens_blur.py.

Example workflow

Depth Lens Blur example workflow estimates depth with the built-in Depth Anything, focuses on the face from Face Detect Lite and blurs the arm in the foreground; crops show the arm before and after. It uses only core ComfyUI nodes and this pack. Copy workflows/sample_portrait.jpg into ComfyUI/input/.

Result: Depth Lens Blur example workflow