ComfyUI Photoshop: bringing a local diffusion backend into an existing editor

A bridge, not a model

The important thing to understand about this repository is that it contains no diffusion model and no inference code. It is a bridge: a Photoshop panel that talks to a ComfyUI server you are already running, sends it work, and brings results back as layers.

That architecture explains the whole shape of the project. The Python side depends on msgpack and aiofiles for the ComfyUI protocol, plus GitPython and PyGithub for the in-plugin update mechanism, with numpy alongside. The panel itself is UXP, Adobe's extension platform, which is why installation goes through the ZXP UXP Installer rather than a normal extension manager.

The repository description puts it plainly: ComfyUI inside of your Photoshop. The README frames the value proposition as putting ComfyUI's generation capabilities into a Photoshop workflow, and the two applications keep their own strengths: ComfyUI supplies the graph and the sampler, Photoshop supplies layers, masks, selections and everything else you already do by hand.

What version 1.9 added

The README carries a patch notes table covering versions 1.0 through 1.9, which is more release history than most projects of this size publish anywhere. Version 1.9 is the current headline and its entries are specific: crop over selection with padding, port change support, tiny shortcuts, a preserve selection feature, an improved Play on Channels behaviour, and convenient in-plugin updates.

Port change support matters more than it sounds. ComfyUI defaults to port 8188, but that is not guaranteed on every install, and a panel that hard-codes the port breaks the first time someone runs a second instance or a non-default configuration. Tiny shortcuts is a vaguer entry, but in a panel where every action is a click in a side dock, fewer clicks is the whole game.

Earlier entries in the table tell you what the project was fighting. Versions 1.4 to 1.6 cover real-time workflow sync, a simpler workflow, enhanced image saving, mask preview, a plugin install button, faster start-up, macOS support and Photopea integration. Version 1.8 added a randomization feature, resizable text fields and panel animations. The progression from 1.0 to 1.9 reads as a project that spent its early life on start-up time and ergonomics before turning to feature work.

Installing the panel and the workflow are separate jobs

The stated minimum is 6GB of VRAM, 12GB of RAM and Photoshop 2022 or newer. Those numbers reflect what ComfyUI needs rather than what the panel needs, since the panel itself is lightweight.

Installation splits into two halves that are easy to confuse. The Photoshop half is a signed extension package installed with Adobe's ZXP UXP Installer. The ComfyUI half is a workflow file that you drop into your ComfyUI directory and then use ComfyUI Manager to resolve whatever custom nodes it references.

Neither half is self-contained, which is the main operational difference from a Photoshop-native AI plugin. The panel cannot start ComfyUI for you, and the workflow cannot be generated without models you already have on disk. There is a note for newcomers that points at a video walkthrough of the ComfyUI installation, which tells you the plugin assumes a working ComfyUI as a precondition rather than bundling one.

The checkpoint list is the actual dependency spec

Most plugin documentation stops at the install button. This README goes further and names every model the workflow expects, which is the part that actually determines whether your first generation works.

Two checkpoints go into the ComfyUI models checkpoints folder: EpicRealism Natural Sin RC1 VAE as the default, and EpicRealism pure Evolution V5 for in-painting. A detailer LoRA goes into the LoRAs folder. Everything else is installed through ComfyUI Manager's model installer: LCM LoRA for SD1.5, three ControlNet v1.1 variants in lineart, scribble and inpaint flavors at fp16, and 4x-UltraSharp.

That list tells you a lot about how the workflow is built. The presence of three ControlNets alongside lineart and scribble variants is why the plugin works for the thing people actually want from it, keeping a rough sketch or edge map rather than generating freeform. The LoRA-based upscaler rather than a native upscale model explains the 4x-UltraSharp entry. LCM LoRA on SD1.5 is what makes the in-panel generation feel fast enough to iterate on.

The README also mentions added language support for Chinese, Japanese and Korean, with a Load SD1.5 button that loads the workflow in whichever language your ComfyUI uses, and compatibility with the AIGODLIKE ComfyUI Translation node.

Reading the repository layout as a Python ComfyUI node

The tree is small and tells you what kind of package this is. There is an `__init__.py` at the root, directories named `js/` and `py/` for the panel and the server-side logic, an `Install_Plugin/` directory holding the distributable extension, a `data/` directory for assets, and a pyproject.toml.

That pyproject.toml matters because it registers the project with the Comfy Registry. It declares a publisher id, a display name and an icon path, which is how a custom node becomes installable through a registry rather than only by copying files. It also records the version:

[project]
name = "comfyui-photoshop"
description = "🔹 Powerfull bridge to Photoshop by NimaNzrii"
version = "1.9.3"
license = {file = "LICENSE"}
dependencies = ["msgpack", "aiofiles", "GitPython", "PyGithub", "asyncio", "numpy"]

So the panel version at install time is 1.9.3, one patch above the 1.9 the README headlines, and the registry metadata is what a ComfyUI Manager install will pull.

The license file name does not match what the manifest expects

This is worth spelling out because it is the kind of detail that quietly matters. GitHub reports the repository license as NOASSERTION, which means its detection found nothing it could map to a standard license. The README links to a file called License in the repository root, and the tree does contain an entry named License, without an extension.

The pyproject.toml, however, declares `license = {file = "LICENSE"}` in uppercase. On a case-sensitive filesystem those are two different names, so the packaging metadata points at a file that does not exist under that exact name.

Both facts are true at once and neither is necessarily an error: git repositories on Windows and macOS are usually case-insensitive, so a checkout there will happily resolve either spelling, while a Linux checkout would not. The practical advice is unchanged either way. Open the License file at the root of the repository and read what it actually says, because neither the GitHub metadata nor the packaging field will tell you reliably.

There are no GitHub releases at all for this project, which is worth knowing given the README documents nine versions. The patch notes table is the changelog. The repository is not archived, it was last pushed on 2026-08-19, and it carries 173 open issues against roughly 1,750 stars, so expect a busy tracker. The README also promotes a successor project, Reezee AI, twice, once near the top and once in the footer, which is a fairly clear signal about where the author's attention is going.

Editorial conclusion

This plugin is a good fit if you already run ComfyUI and want its workflows reachable from inside Photoshop instead of round-tripping files through a browser tab. The panel handles selection transfer, layer compositing and a set of quality shortcuts that are tedious to rebuild by hand. The costs are equally clear: no GitHub releases exist despite a nine-version history in the README, the license file name does not match the name the manifest expects, and the author has already pointed readers at a successor project. Treat the README's version table as the changelog, verify the panel version against the installed build after an in-plugin update, and read the License file directly rather than relying on the repository metadata.

Frequently asked questions

Official sources

1. Issues
2. NimaNzrii/comfyui-photoshop on GitHub
3. README