Photoshop Live Preview

The output side of the realtime loop, and the reason the loop stays in Photoshop at all. Wire your rendered IMAGE into Photoshop Live Preview and each result appears in the ComfyUI Preview panel docked beside your canvas - right where you're drawing. You see your sketch, you see the AI's take on it, you don't look at a browser tab. It's the "live" in the live loop made visible.

What it does behind the scenes

The node takes a single image input and has no outputs - it's an output node, an endpoint. Two things worth knowing:

• The panel only displays a small copy of the render, but the node keeps every render at full quality server-side. That's not just vanity: it's what makes the refine pass and Add-as-a-layer work. When you hover a render in the panel you get Add as a layer (drops it into your document, scaled to the canvas) and Refine (runs the refine branch) - and both of those place the full-quality version, not the thumbnail.
• Without the plugin connected, the node quietly does nothing rather than failing your render. It runs, logs a no-op, and your image stays where it already was - in ComfyUI. That's a deliberate design choice across this pack: anything that can't deliver somewhere still doesn't break the graph.

Where it sits in the loop

It's the terminal for the four-node draw loop: Live Canvas (drawing in) → KSampler → Live Preview (render out). For the optional high-quality step, pair it with Photoshop Refine Source and Photoshop Add Layer: Refine runs once at high quality, and the result lands either back in the preview pane, into the document as a layer, or both, depending on how you wire it.

Install

Same as every node in this pack - Manager → Install via Git URL, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ericpaulsnowden/comfyui-photoshop-bridge.git comfyui-photoshop-bridge
pip install -r comfyui-photoshop-bridge/requirements.txt

…with ComfyUI's own Python, then restart - and add the Tier 2 Photoshop plugin, because without it this node has no wall to throw the image at. Pre-1.0 and actively developed. If the panel shows renders but they lag your strokes, that's the few-step-model problem surfacing again: the example loop is tuned for an LCM/Turbo-class model, and a plain checkpoint at 4 steps will produce something that looks like your drawing with a prompt that doesn't stick.