PS Blend

Sooner or later you hit a composite job that plain "add" or "multiply" won't cut. Maybe you're blending a relit subject back into a scene after an IC-Light pass, feathering an inpainted region so the seam disappears, or baking a texture onto something and wanting it to actually bite into the surface. ComfyUI's built-in blending is thin, so you either install a kitchen-sink toolkit or grab this: one node, PSBlendNode, that ports Photoshop's blend modes into ComfyUI as plain tensor math.

That's the whole pack, by the way. No API, no key, no model download, nothing to fine-tune. It's a single node that takes two images, a mode, and an opacity, and returns one composited image. That focus is the selling point - if WAS Node Suite's 400-node post-processing arsenal feels like overkill for a blend step, this is the lean alternative.

How it works

Every mode is a textbook Photoshop/W3C blend formula written in PyTorch. On each run the node picks a device (CUDA, then Apple MPS, then CPU), checks that both inputs have identical dimensions, applies the formula channel-by-channel, and mixes the result toward the base image by your opacity. No neural network anywhere - it's deterministic math, so it's fast and it works on any hardware that runs ComfyUI.

One detail worth knowing: the code actually separates and blends alpha channels, even though the README claims it doesn't support transparency. On ordinary 3-channel images the alpha slice comes back empty and the step silently does nothing, so nothing breaks - just don't expect meaningful alpha handling with RGBA inputs.

The inputs that matter

There are only four, which is the point:

• image1 / image2 - two IMAGE tensors. Both must be the same size; the node raises a ValueError if they're not, so resize or crop first.
• blend_mode - the dropdown, and it's a full 27 entries: Normal, Dissolve, Darken, Multiply, Color Burn, Linear Burn, Darker Color, Lighten, Screen, Color Dodge, Linear Dodge, Lighter Color, Overlay, Soft Light, Hard Light, Vivid Light, Linear Light, Pin Light, Hard Mix, Difference, Exclusion, Subtract, Divide, Hue, Saturation, Color, and Luminosity. That's the whole Photoshop roster, including the HSL group (Hue/Saturation/Color/Luminosity) that most quick blend nodes skip.
• opacity - a float from 0 to 1, default 1.0. Pull it down to ease between the original base and the full blend, just like Photoshop's layer opacity slider.

The output is a single IMAGE. Wire it into a PreviewImage node to look, a VAE Decode to save, or straight into whatever's next in your pipeline - an upscale pass, an img2img step, a controlnet-conditioned pass.

Installing it

Easiest path is ComfyUI Manager - search for ComfyUI_PS_Blend_Node (or "PS Blend") and hit Install. Or clone it by hand:

cd path/to/ComfyUI/custom_nodes
git clone https://github.com/bluevisor/ComfyUI_PS_Blend_Node.git PSBlendNode

Then restart ComfyUI. The only declared dependency is torch, which ComfyUI already ships, so there's nothing else to download - that's the whole install story, and it's why this node is so painless.

Gotchas

• Same-size or nothing. Mismatched images throw ValueError: Images must have the same dimensions. Slap an ImageScale or crop node on the smaller one first.
• Input order isn't consistent. For symmetric modes (Multiply, Screen, Overlay) it doesn't matter, but the directional ones - Subtract, Divide, the Dodge/Burn family - don't agree on which input is "base" and which is "blend" across modes. If a result looks inverted, just swap image1 and image2; it's a two-second fix.
• Dissolve is random. It's built on a per-run random mask, so the same inputs give you different noise each execution. Don't build a reproducible pass around it.
• No alpha out. As covered, transparency support is cosmetic at best - keep your masks on separate nodes.

It won't change your life, but when you need a Multiply or a Soft Light that behaves exactly like Photoshop, this is the one-node answer.