🔹Photoshop ComfyUI Plugin

If you live in Photoshop and keep tabbing over to ComfyUI every time you actually want to generate, this is the node that closes the loop. "🔹Photoshop ComfyUI Plugin" is the input side of the comfyui-photoshop bridge - it pulls your live Photoshop document straight into the graph, so your canvas, your selection, and even your prompt can all stay in Photoshop while ComfyUI does the heavy lifting.

What it actually is

This is the node the whole pack is built around. It has zero inputs and eight outputs - it's a source node, a way to get Photoshop's current state into the workflow:

• Canvas (IMAGE) - your document, as it sits in Photoshop right now.
• Mask (MASK) - your selection, as a mask.
• Slider (FLOAT) - the plugin's strength/denoise dial, normalized to 0–1.
• Seed (INT), "+" / "-" (STRING) - the plugin's seed and positive/negative prompt fields.
• W / H (INT) - canvas dimensions, handy when you're sizing latents.

The pitch was never subtle. The author demoed real-time diffusion inside Photoshop back in late 2023 - "no OBS, no virtual cam" - and the reception was loud. If you tried the Krita AI Diffusion route and bounced because you live in Adobe, this is the same idea for your editor of choice.

How it works

It is not screen capture, and it doesn't call any external API. The Photoshop plugin (a UXP extension, installed as a .ccx) opens a websocket to ComfyUI's own server at /ps/ws. When you paint, it pushes your canvas, your mask, and a small config JSON as base64. The pack's Backend writes those into data/ps_inputs/ - PS_canvas.png, PS_mask.png, config.json - and this node reads them back every run, converts the PNG to a float tensor, pulls the blue channel into a MASK, and surfaces the config as those +/–/Seed/Slider outputs.

Two details worth knowing. First, the mask is your selection: select a region in Photoshop and that's exactly what the node hands you - perfect for feeding an inpainting setup without ever opening a mask editor. Second, the node only re-runs when the files actually change (it hashes them), so a live loop isn't churning the graph every frame for nothing.

Getting it running

Two halves, and both have to work. ComfyUI side first:

cd ComfyUI/custom_nodes
git clone https://github.com/NimaNzrii/comfyui-photoshop

or hit ComfyUI Manager → Install Custom Nodes → search comfyui-photoshop → install, then restart. The pack's own deps (msgpack, aiofiles, GitPython, PyGithub, numpy) install with it; it's not a heavy pack.

The Photoshop side is the fiddly bit. The README's path: download the .ccx and install it with the ZXP UXP Installer, restart Photoshop, then launch the plugin from the Plugins menu. (There's also a bundled installer.py that copies it into the right UXP folder for you.) Then drop in the reference workflow linked in the README, install its missing nodes via Manager, and let Manager's "Install Models" pull the checkpoints: EpicRealism Natural Sin RC1 plus the inpainting variant, the Detailer LoRA, an LCM LoRA for speed, the SD 1.5 ControlNets (lineart, scribble, inpaint), and 4x-UltraSharp. Minimum spec per the README: 6GB VRAM, 12GB RAM, Photoshop 2022+.

Where people get burned

The .ccx install is the classic wall - the ZXP installer sometimes refuses the file outright, and the Creative Cloud marketplace route fails with network errors. The fix that kept coming up in the community: do a clean install of Portable ComfyUI, because bundled environments like Stability Matrix could break the backend bridge entirely.

No live update? Check the backend, not the node. The plugin and ComfyUI talk over localhost websockets, so they must run on the same machine and the plugin has to be launched - watch for the # PS: Photoshop Connected line in ComfyUI's console.

One last thing: this node is specifically the UXP-plugin path. If you'd rather skip installing a Photoshop plugin at all, the pack's separate "🔹 Photoshop RemoteConnection" node is the socket-based alternative - different tradeoffs, on its own page.