EmAySee Load Image + Name

LoadImage, plus the thing you always wanted from it

Core ComfyUI's LoadImage gives you an image and a mask - that's it. The moment you build anything dataset-shaped, you discover you also need the filename in the graph: to write a matching .txt caption, to log which image failed, to tell an LLM "this file is portrait_0042.png" so its captions stay aligned with your files. EmAySee_LoadImagePlusName is a drop-in clone of the standard loader with one extra output: filename_stem, the file's basename minus extension.

It's from the EmAySee pack, and for anyone captioning or tagging a folder of images this is the node that makes the loop close: load image → LLM describes it (see EmAySee_LlamaVision) → save the text with the filename from filename_stem → advance. That's a real workflow, not a hypothetical - it's the exact shape of a training-dataset prep pass, which the LoRA-training world lives on.

How it works

Same code path as the core loader: it reads the file from ComfyUI's input directory, applies EXIF orientation transpose, converts to RGB, and normalizes to a float tensor. The alpha channel, if present, becomes the mask output (inverted, matching ComfyUI's convention that white = masked region). No alpha? You get a flat 64×64 zero mask as a placeholder. filename_stem is computed with os.path.splitext on the selected file, so cat.png → cat.

Inputs and outputs

• image - a dropdown of files in your ComfyUI/input/ folder, with the upload button enabled.

Outputs:

• image (IMAGE) - the loaded RGB tensor.
• mask (MASK) - the alpha channel as a mask, or a 64×64 zero placeholder when there's no transparency.
• filename_stem (STRING) - the name without extension.

Install

From ComfyUI_EmAySee_CustomNodes:

cd ComfyUI/custom_nodes
git clone https://github.com/EmAySee/ComfyUI_EmAySee_CustomNodes
# restart ComfyUI

Or via ComfyUI Manager → "ComfyUI_EmAySee_CustomNodes". No extra dependencies or downloads.

Gotchas

The mask is the trap. That 64×64 zeros placeholder when the image has no alpha is not a real mask - it's an empty tensor that will silently do nothing (or worse) if you feed it into a mask-gated pipeline expecting coverage. Guard it with the pack's EmAySee_MaskGuard or check your workflow assumptions before relying on it. Also note the dropdown only lists the input folder - same as core LoadImage, so subfolders need path selection. And it's a plain loader: no batching, no automatic next-file advance, so driving a whole folder means looping the workflow, not pointing the node at a directory.