近乎零成本的 AI 藝術：為何元資料、封裝與追蹤只需 Python

Last March I lost a full weekend to a CSV header.

I had 40 assets ready for Adobe Stock. I wrote the upload file by hand in a spreadsheet, hit submit, and the whole batch bounced back in under a minute. The reason: my metadata file had the columns in the wrong order. Adobe's uploader checks the header row before it checks anything else, and if a single column name or position is off, it rejects the file instantly. No partial credit. Every asset, gone from the queue, and I got to find the mistake by reading a forum thread from 2019.

That was the moment I stopped treating the art as the project. The art was done. The project was everything around the art, and almost none of it costs money. It costs Python.

Metadata is a formatting problem, not a creative one

Each stock platform has its own required CSV shape, and they do not negotiate. Adobe Stock wants a 5-column header. Vecteezy wants 4 columns. Dreamstime wants 15. Get the header wrong and the upload dies before a single image is reviewed. Get it right and the same folder of files can feed all three channels.

So I wrote make_metadata.py. It walks a folder, reads the filenames, and emits the exact header each platform demands. Adobe mode writes 5 columns, Vecteezy mode writes 4, Dreamstime mode writes 15. Nothing else. It uses only the Python standard library, so there is nothing to install and nothing to break. If you pass a titles file it uses your real titles and keywords. If you don't, it title-cases the filename, which is fine for a first pass but you should replace it before you submit, because reviewers do read titles.

The whole "cost" of solving metadata across three platforms was one afternoon and zero dollars.

Double-posting is the fastest way to get flagged

Here is a failure mode nobody warns you about. You run a batch upload at 2am. Something looks wrong, so you fix one thing and run it again. Now the same asset exists twice on your account, and repeated duplicates are exactly the pattern that gets contributor accounts reviewed and flagged.

I did this once with a sticker pack. Two copies of every file. Cleaning it up took longer than making the pack.

The fix is boring: a ledger. upload_tracker.py keeps a plain record of what you have already posted. pending shows what is left, done marks an item, status gives you the picture. A re-run can never double-post because the tracker remembers. Standard library only, no install. The unglamorous scripts are the ones that protect your account.

Packaging is where a file becomes a product

A bare file does not sell. A zip with a real cover does, and the difference is one command.

pack_product.py takes a folder, zips it, and renders a cover at 1280x720 plus a thumbnail at 600x600. Those two sizes cover what most marketplaces and social posts actually ask for. It writes everything into a _system folder, and you can redirect that with --dir or an environment variable. The only dependency is Pillow.

For docs, md2pdf.py turns a Markdown file into a styled PDF with headings, tables, lists and links, using reportlab. One command in, a product-shaped PDF out.

Across my catalogue these two scripts packaged 39 finished products, about 856 MB of output, on a server with 1 vCPU and 2 GB of RAM and no GPU.

The tiny-server habits that made it possible

Running image work on a 1-core, 2 GB box teaches you discipline fast. The habits that matter:

One process at a time. Never fan out image jobs in parallel on a single core, you just thrash.

Free memory between items. A gc.collect() after each asset keeps the pipeline from creeping toward the ceiling.

Measure peak RSS, not average. resource.getrusage(RUSAGE_SELF).ru_maxrss tells you the real number that will kill your process.

One file per subprocess for heavy render steps, so memory actually returns to the OS.

Render at the resolution you need. Inspecting a preview at 60 dpi instead of 300 is the difference between fitting in RAM and not.

And one habit about folders: keep a fixed layout, assets, output, packs, research, scripts. Keep research/ forever. I deleted my notes folder once to "clean up" and regretted it within a month. Your notes and logs are the raw material for the next product.

What the whole thing actually costs

Add it up. The metadata scripts: free, standard library. The tracker: free. Packaging: Pillow, free. PDF: reportlab, free. Two third-party packages total for the entire layer. Server: 1 vCPU, 2 GB, the cheapest tier that exists. The expensive part of this business, if there is one, is the image generation itself, and even that is a per-image cost you control by batching sensibly, not a subscription you are locked into.

The part that stalled me was never the art. It was the four unglamorous layers around it: metadata, tracking, docs, packaging. Once those were scripts, one prompt engine turned into 39 finished products.

I put those exact scripts together as the Pipeline Starter Kit. Four standalone Python files, no framework, no cloud account, the same ones that ran on the 2 GB box. If you would rather not lose a weekend to a CSV header, it is here:

See the kit

Building digital products? Free guides & kits for digital product sellers.