图像水印

If you post generations anywhere, your options have historically been a batch action in an image editor or a node that sort of works. 图像水印 (QG_ImageWatermark) is the "sort of works" problem solved: paste one image onto another, positioned by anchor plus offset, scaled four different ways, rotatable, fadeable, maskable, and applied to every frame of a batch.

It's not glamorous and it's the kind of node you install for one job. But the fact that it happens inside the graph means the watermark is present in the file that gets saved, including on automated runs where nobody's opening Photoshop.

How it works

Two inputs matter: base_image (the picture) and watermark_image (the logo). The watermark gets scaled according to scale_mode:

• 不缩放 - leave it at native size. Fine if you generated it at the right dimensions.
• 等比适配 - shrink to fit inside the base image, aspect preserved. For a big logo on a small render.
• 按倍数缩放 - multiply native size by scale_factor (0.01–16).
• 按指定宽高缩放 - force target_width × target_height, ignoring aspect. This one will squish your logo if the numbers aren't proportional; that's on purpose, sometimes you want exactly 512×64.

Then position picks one of nine anchors - 居中, 上, 下, 左, 右, and the four corners - and offset_x / offset_y nudge from there in pixels. Both default to 32, which is a sane bottom-right corner inset, and offsets can be negative (range is ±48000, so no realistic excuse to be limited).

rotation is an integer in degrees, clockwise, and the canvas expands around the watermark rather than clipping the corners - the pack notes this as a bug fix over the version it was ported from. opacity is the field to read the tooltip on: 0 is fully opaque, 100 is fully transparent. That's inverted from intuition and from most software, and it's the thing people get wrong first. A value of 25 means you're seeing three quarters of your logo.

The optional mask input is subtler than it looks. White in the mask means hide the watermark, black means show it. And the mask is resized to the watermark's dimensions, not the base image's - so it's a cutout for the logo itself, not a keep-out zone on the picture. Use it to knock out a portion of the logo (a shape you want it to sit behind, a soft fade), not to restrict where on the image the watermark lands.

Batches work the way you'd hope: base_image applies the watermark to every frame, watermark_image uses only its first frame, and the mask uses its first frame too.

Inputs and outputs

Required, in practice: base_image, watermark_image, scale_mode, position, offset_x, offset_y, rotation, opacity, plus scale_method, scale_factor, target_width, target_height for the scaling modes. Optional: mask. Output: IMAGE, same frame count as your base.

Wire the output into SaveImage and you're done, or into a 图像对比 node first to check the placement at full size rather than squinting at a node preview.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/live1617/ComfyUI-QG-tools

Full restart; console prints [QG-tools] v0.9.2 加载完成 | 节点数 24. Manager search for ComfyUI-QG-Tools covers the Registry case. Nothing to pip install - the requirements file is deliberately all comments so Manager skips pip entirely.

Ported from ComfyUI_Muye (MIT), searchable as watermark, logo, stamp or 水印.

Where people get burned

Opacity inverted, again. 0 = solid. If your watermark seems to be missing, check whether it's sitting at 50 thinking you asked for half strength - you did, you're getting half. If it's invisible, it's near 100.

The watermark has to be an IMAGE, not a path. There's no file picker here; feed it through a LoadImage. A common build is LoadImage → watermark → done, and it's worth grouping that pair so you don't accidentally delete the loader.

Transparent PNGs used to be broken. The pack's changelog calls this out explicitly: upstream's paste ignored the alpha channel, so a logo with transparency got pasted as an opaque rectangle. It's fixed here, and rotated corners now composite correctly. If you're comparing against the original node in another pack, that's the difference.

Rotation and offsets interact. Rotation happens before the anchor placement, so a 45° rotated logo whose bounding box is now bigger may poke outside the base image. The canvas expands to keep the watermark whole, not to keep it on the picture - offset it inward or scale it down.