Load Image (from Outputs)
Pick up where your last run left off
- IMAGE
- MASK
The classic ComfyUI workflow is one long graph: generate, upscale, refine, save, repeat. But a huge fraction of people run it the smarter way - generate first, cull the trash, then run a separate refinement workflow only on the keepers. Load Image (from Outputs) is the node that makes that second workflow painless, because instead of digging through your file system, it hands you the images ComfyUI itself just produced.
What it is
It's Load Image with the file picker pointed at your output directory instead of your input directory. Same image dropdown, same outputs - IMAGE and MASK - same underlying loading code (it literally subclasses Load Image). What's different is the widget behavior, which is where the real value lives: the dropdown has a refresh button, and the node is designed to auto-select the first (newest) output whenever you click it. The node's own description says it plainly: click refresh and it updates the list and selects the first image, "allowing for easy iteration."
So the loop becomes: run your generation workflow, look at what came out in the output folder, delete the bad ones, then in your refinement workflow hit refresh on Load Image (from Outputs) and queue. It pulls the newest surviving output without you touching the file dialog. For anyone doing the "upscale/refine only the good ones" dance - a workflow the community describes constantly - this is the built-in version of that.
What it outputs
IMAGE- the image tensor, ready to feed img2img, upscalers, FaceDetailer, or whatever your refinement graph does.MASK- derived from the image's alpha channel, same convention as Load Image. Most PNGs have no alpha, in which case it's an all-zeros mask (i.e., "nothing to regenerate"); the mask output is there for consistency with Load Image rather than being useful on its own.
The details that matter
The node is marked experimental in core, and it was added precisely for the iterate-on-your-own-outputs pattern. It reads files from the output directory via the /internal/files/output route - the same folder your Save Image node writes to, including any subfolders you've created with folder-style prefixes. If you save with folder1/name prefixes, your outputs live in output/folder1/ and they show up here too.
Where people get tripped up: "newest" is determined by the output listing, so if you've been running a workflow that writes a batch, the refresh picks the first file in the refreshed list - which is usually the latest run, but sorting assumptions can bite if you have multiple workflows writing to the same output folder with different prefixes. And the mask output being zeros can confuse anyone who wires it straight into inpainting expecting their saved alpha - this node is about loading results, not masks.
Getting it
Ships with ComfyUI core, a newer experimental addition - no install, no model files. It's one of those nodes you won't need until you adopt the two-phase workflow, and then it quietly becomes load-bearing. It's also the built-in answer to the pattern people used to bolt on with WAS Suite's incrementing batch loader: quick, single-image iteration straight from your own outputs.
Inputs (1)
| Name | Type | Default | Description |
|---|---|---|---|
| image | COMBO | — |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| IMAGE | IMAGE | — |
| MASK | MASK | — |