What to do with thousands of AI-generated images

A practical Mac workflow · Updated July 25, 2026

Generating images is so cheap that everyone who runs Stable Diffusion, ComfyUI, or Draw Things eventually hits the same wall: an outputs folder with thousands of files, no idea what's in there anymore, and a vague guilt about the disk space. This is a practical workflow for digging out — and staying out.

Why the pile happens

The tools are built to generate, not to organize. Automatic1111 drops everything into dated subfolders under outputs/; ComfyUI numbers files into one flat output/ folder; Draw Things keeps a library you export from. None of them asks whether an image was any good — the misses land next to the keepers, forever, in filename order.

The result is a folder where the only way to find last month's good batch is to scrub through everything around it. The fix isn't a better viewer; it's a one-time triage plus a small habit.

Step 1 — Decide what "keep" actually means

Three buckets cover almost everyone:

Keepers — the images you'd post, print, use in a project, or train from. For most people this is well under 5% of a batch.

Reference — not portfolio material, but the settings matter: a composition worth rerunning, a style test, a seed you'll want again. Here's the useful part: for images from Automatic1111, Forge, ComfyUI, and Draw Things, the recipe travels inside the file — prompt, negative prompt, seed, model, sampler, all embedded in the PNG or JPEG. A reference image is a saved recipe.

Everything else — misses, near-duplicates, six-fingered hands. This is the bucket that makes triage feel safe: because your keepers and references carry their own generation settings, deleting the rest costs you almost nothing. You're not losing work; you're deleting the takes that didn't land.

Step 2 — A folder structure that scales

Organize by project, not by tool or date — the tools already did date-and-tool, and that's exactly the structure that stopped working. Something like:

AI Art / ProjectName / keepers
AI Art / ProjectName / reference
AI Art / _style-tests

Two rules keep it honest. First, the outputs folder is an inbox, not an archive — things leave it, they don't live there. Second, don't over-nest: if filing an image takes more than one decision, you'll stop filing.

Step 3 — The triage pass

Go through the backlog in sittings, newest first — recent batches are the ones you actually remember and use. Sweep the obvious rejects in bulk first (any grid view where you can select a run of misses and delete them together), then make the keep/reference calls one image at a time, big enough to actually judge.

The pace matters more than the method: a keep-or-toss call should take about a second. If you're hesitating, that's what a Maybe folder is for — decide once, move on, and let the second pass through Maybe be short.

The tools for the job (honestly)

Finder + Quick Look works for a hundred images: arrows to move, Space to peek, drag to file. It breaks down at a thousand — every move is a drag, nothing shows you the prompt, and one mis-drop into the wrong folder is silent.

Your generator's own browser — the A1111 image-browser extension, ComfyUI gallery nodes — shows metadata but lives inside the web UI, only sees that tool's outputs, and doesn't file anything into your real folders.

Diffusion Toolkit is the mature library manager for this — scan, search, tag by metadata — but it's Windows-only.

FileSift is our tool, and the Mac-native way to do the triage pass itself: each image shows big with its embedded prompt, seed, and model beside it, one number key files it (or trashes it), and ⌘Z physically moves it back if you change your mind. It reads Automatic1111, Forge, ComfyUI (including Flux), and Draw Things metadata; Midjourney and ChatGPT images sort just as fast but don't carry a readable prompt to show. Everything happens locally — nothing is uploaded.

Step 4 — Keep it from happening again

The habit that sticks: triage per session, while you still remember why you generated the batch. End of a prompting session, five minutes, outputs folder back to empty. If you keep destination folders consistent (a saved preset helps), the decision cost drops to almost nothing — and the pile never comes back.

Back up the keepers and reference folders like real work, because they are. The outputs inbox needs no backup: anything still in it is, by definition, undecided.

The short version

Keepers, reference, gone — three buckets. File by project. Sweep rejects in bulk, judge the rest one at a time with the metadata visible. Triage each session while it's fresh. Your keepers carry their own recipes, so delete without fear.

FileSift is a native Mac app for exactly this — see how it works on AI images, or find it on the Mac App Store. Requires macOS 14 or later.