萌宝AI·保存图片

Most "improved Save Image" nodes are a trap: they reimplement saving, get the metadata wrong, and quietly break the one thing that makes ComfyUI images shareable - the embedded workflow. MengBaoSaveImage (萌宝AI·保存图片) does the opposite. Look at the source and it literally reaches into ComfyUI's own nodes module, deep-copies the native SaveImage input schema and calls the native implementation. Same output/ directory, same filename_prefix formatting with %date:...% and %Empty Latent Image.width% style tokens, same batch handling, same alpha, same PNG workflow metadata. If the native node changes, this one follows.

So what does it add? A display panel and two social features.

The panel, and the prompt library

Instead of the tiny native preview strip, you get an image area with batch switching and two buttons: Add to material library and Add to prompt library. The first stores the currently displayed image as a proper library asset - original dimensions and alpha preserved, validated against the library's size rules, never a thumbnail standing in for the original. That's the same store that MengBaoMaterialLibrary reads, so a good generation becomes a reusable reference in one click instead of a folder expedition.

The second one is the interesting one. Click it and the prompt-library form opens pre-filled with a 512px thumbnail of the image and the prompt that produced it. You get to edit the title, group, text and tags, then save. Since ComfyUI's own PNG metadata already embeds the workflow, the point here isn't recovery - it's having a searchable, human-named index of prompts you'd otherwise dig out of files forever.

Where does the prompt come from? The node walks the graph backwards from its images input looking for a recognised source: it stops at the pack's own generation node (WANGImageAPI) and reads its prompt, or it finds a KSampler/KSamplerAdvanced/SamplerCustom and walks its positive input to a CLIPTextEncode to read the text. Negative prompts and unrelated reference branches are explicitly excluded. If several candidates turn up, you pick one in the panel.

If your prompt comes from anything it doesn't recognise - a third-party text node, a dynamic widget, the pack's prompt reader - the optional generation_prompt input is the answer: anything you wire there outranks the auto-detection. Leave it empty and the node may open a blank draft rather than guess wrong, which is the right call even though it feels like a missing feature the first time.

Inputs and outputs

images (IMAGE) and filename_prefix (STRING, default ComfyUI, with the native tooltip about formatting tokens), plus the optional generation_prompt string input. That's it - no seed, no metadata toggles. Nothing comes out of the node in graph terms: it's a terminal output node, so it executes and that's the end of the branch.

Installing it

17 nodes, one pack: ComfyUI-MengBaoAI, Corkery520, MIT, Registry id mengbaoai. Manager search: MengBaoAI, mengbaoai, 萌宝AI. CLI: comfy node install mengbaoai. Git while the Registry listing waits on its first release:

cd ComfyUI/custom_nodes
git clone https://github.com/Corkery520/ComfyUI-MengBaoAI.git
cd ComfyUI-MengBaoAI
python -m pip install -r requirements.txt

numpy, Pillow, requests - and one requirement the README states outright: import the native SaveImage node normally, because this node delegates to it and will raise if it's missing. After installing, Ctrl+F5 so the panel's JavaScript loads; the pack's docs treat that as routine after any update, since the buttons call local HTTP routes.

Things that bite

• Library saves live in ComfyUI/user/mengbaoai/, not in output. Back that folder up or the index is gone, even if the PNGs are still on disk.
• The panel is a result view, not an upload target - you can't drag an image into it to save. Generate or reroute the pixels.
• Auto-detected prompts can be wrong when your graph is clever. Two generation nodes upstream, a prompt assembled by concatenation, a reroute-heavy layout: wire generation_prompt explicitly and stop guessing.
• Nothing here replaces --disable-metadata. If your ComfyUI is configured not to write prompt metadata into PNGs, this node still won't - it inherits that setting because it is the native node.