图南PS发送器

If TuNanPSBridge is the in-box, TuNanPSSender is the out-box. It takes the finished image from your graph and pushes it back into the same Photoshop document it came from - positioned exactly where you selected. That's what makes the whole loop feel less like "two apps shouting over a socket" and more like Photoshop with a genuinely better img2img. It's the last node in the workflow, and it's a dead end on purpose: it has no outputs. Wire your VAEDecode into it and stop.

How it works

Like its sibling, it runs on the same localhost HTTP/WebSocket connection your Photoshop plugin (.ccx) opened. On execution it grabs the placement snapshot and selection mask the bridge captured, encodes your image as PNG, and ships it back. Two paste modes decide where it lands:

• 选区还原模式 (selection restore, the default): composites the result back into the original selection bounds, so the generated pixels land precisely where you marquee'd in Photoshop. This is the mode that makes inpainting-in-PS feel magical.
• 整图模式 (whole image): returns the full frame as-is, no positioning.

If Photoshop isn't connected when the run finishes, the node doesn't crash - it renders a preview, keeps it, and reports that the result never made it back. You'll see "waiting"/"preview only" status instead of a delivered image.

The inputs that matter

Only one wire, really: 图像 (IMAGE), straight off VAEDecode. Everything else is tuning:

• 回贴模式: selection restore or whole image. If PS has no live selection when you run, selection restore quietly falls back to whole-image and tells you so - expected behavior, not a bug.
• 边缘收缩 (0–128): erodes the paste mask by a few pixels before compositing. Its job is killing the fringe of unchanged pixels that hugs a hard selection edge - the "white halo" look.
• 边缘柔化 (0–128): feathers the paste edge to hide the seam. Both edge controls only bite in selection-restore mode; in whole-image mode they're no-ops.
• 图像名称 (optional): names the layer/document the result lands on. Defaults to 图南画桥.

Getting it installed

Same pack, same steps as the bridge node: ComfyUI Manager (search tunan-paint-bridge-nodes or 图南画桥), or:

cd ComfyUI/custom_nodes
git clone https://github.com/beiyue0208ly/tunan-paint-bridge-nodes
cd tunan-paint-bridge-nodes
pip install -r requirements.txt

Restart, then install the matching "图南画桥" Photoshop plugin (.ccx) from tunanart.cn or the releases page. Version-match it to the nodes (1.0.x ↔ 1.0.x). Dependencies are just aiohttp, numpy, Pillow, scipy - no model downloads.

Where people get burned

The number one confusion is the missing output sockets - people spend real minutes hunting for a place to wire the result out. There isn't one; the result goes to Photoshop, not to another node. Second: "I ran it and nothing appeared in PS." That's almost always a connection problem - check the bridge node's status, confirm the plugin is installed and version-matched, and make sure the document you're sending from is still the active one. Third: the edge controls doing nothing in whole-image mode. That's the design, not a malfunction. If you want the seam-blending controls to do anything at all, use selection-restore mode with a live selection.