Comfy-DepthDiff

Comfy-DepthDiff

A single ComfyUI node that turns a source image into a luma/depth-driven Differential Diffusion mask — dark regions get more diffusion, light regions get preserved (or vice versa). Bakes in the model patch, VAE encode, and mask preview.

What it does

Differential Diffusion selectively re-noises regions of a latent based on a grayscale mask. DepthDiff builds that mask from your input image — either directly from luma, or by running Depth Anything V2 first — and applies everything (model patch, latent noise mask) in one node.

Install

Via ComfyUI Manager (recommended): search DepthDiff and click Install.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/spiritform/Comfy-DepthDiff.git

Restart ComfyUI.

Dependencies

Core node has no external Python dependencies — works out of the box.

Optional (only needed for depth_mode):

• Kijai's comfyui-depthanythingv2 must be installed (via ComfyUI Manager or manual clone).
• Depth Anything V2 safetensors weights auto-download to ComfyUI/models/depthanything/ from Kijai/DepthAnythingV2-safetensors on first use.

If depth_mode is off, feed a pre-computed depth or luma map into the image input directly and neither is required.

Inputs

• model — MODEL (patched with Differential Diffusion internally)
• image — IMAGE (encoded to latent internally; also the mask source)
• vae — VAE (used for the internal encode)
• mask — MASK (optional; gates the final mask, blurred by mask_blur)

Widgets:

• depth_mode — if on, runs Depth Anything V2 (Kijai) on the image before extracting luma
• depth_ckpt — which Depth Anything V2 safetensors checkpoint to use
• depth_max_size — caps the depth pass long-side (default 1024)
• invert — flip the mask (default true; dark = more diffusion)
• input_black / input_white — levels remap (0–255)
• gamma, brightness, contrast — tone shaping
• image_blur — edge softness on the depth/luma mask
• mask_blur — edge softness on the optional input mask
• strength — per-pixel clip on the final mask
• diff_diffusion_multiplier — global multiplier on when the mask triggers denoising during sampling (1.0 = default, >1 = more aggressive, <1 = more preservation)

Outputs

• model — patched MODEL → KSampler
• latent — LATENT with noise_mask attached → KSampler
• mask — MASK output for downstream use

Wiring

Load Image ─┐
            ├─► DepthDiff ─► model ─► KSampler
Checkpoint ─┤              ─► latent ─┘
    └── VAE ┘

License

MIT