LoadImageBypass

Core ComfyUI's Load Image always loads. If you want a workflow that sometimes feeds an image and sometimes doesn't - a text-to-video branch vs an image-to-video branch, A/B testing a condition, or just trying things out - you normally end up deleting and re-adding nodes, or faking it with a switch node downstream. LoadImageBypass is a Load Image with an enable toggle built in, and here's the thing that makes it different from every other "load image" you've used: flip it off and the node returns nothing at all, which tells ComfyUI the branch is dead. No placeholder, no blank - a genuine cut.

How it works

The mechanism is embarrassingly simple, which is why it works. When enable is on, it loads the file, converts to RGB, and hands you an IMAGE tensor like any loader. When enable is off, it returns None. ComfyUI treats a None output as "this branch stops here" - anything fed by that wire gets nothing, so the graph downstream effectively goes silent.

Two real differences from the built-in Load Image, both worth knowing before you swap it in:

1. It resizes your pixels. This node snaps every loaded image to a multiple of 32 using Lanczos. The core loader leaves your file untouched. Feed it a 1000×562 source and you get 992×576 back, silently. For most SDXL/SD1.5 work that's actually what you want - latents prefer multiples of 32/64 - but if you were relying on the loader to pass the original resolution through for, say, a hi-res pass, this will surprise you.

2. The file dropdown is filtered. It only lists actual image files (png, jpg, webp, bmp, tiff, gif, apng), so it won't show the stray text files or configs that clutter the core loader's list. Small quality-of-life thing, but nice.

The inputs are just image_file and enable; the single output is image (IMAGE).

Where people get burned

The obvious one: "bypass" here is the author's name, not ComfyUI's right-click bypass feature. Those are different mechanisms - this is a manual toggle, not the graph's built-in bypass. Don't go looking for a link to the built-in feature.

The sharper trap is the None output. If you take a working graph, swap the core Load Image for this one, and toggle it off, whatever you had feeding off that image (a VAE Encode, a CLIP vision node, a reference-image conditioner) will suddenly receive nothing - and some of those nodes just sit there, while others throw errors that are confusing because nothing obviously changed in your graph. If you use the off switch, wire what follows it to tolerate missing input, or keep a bypass node ready to route around it. This node is genuinely handy for optional-image branches - just know what "off" does to everything downstream.

Installing it

It ships in ComfyUI Fictiverse Nodes. Install via ComfyUI Manager (search ComfyUI Fictiverse Nodes) and restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Fictiverse/ComfyUI_Fictiverse

Then restart ComfyUI. No dependencies, no model files to download - the pack is pure Python plus what ComfyUI already includes.