I lost a whole Saturday to this. I had 40 vectors queued for Adobe Stock, ran the batch, went to bed, and woke up to a review queue full of rejections. Same reason on every one: quality issues on the vector exports. Not keyword problems, not category problems. The files themselves.
Here is what I actually did wrong, and what the fixed pipeline looks like now.
The trace was the problem, not the art
My prompt engine was producing clean source images. I checked them at full resolution before tracing. But my autotrace step was set to a default color count and a default noise threshold, and it was doing two things wrong at once.
First, it was merging near-identical colors. On a gradient-heavy illustration this produced banding that looked fine at 100% zoom in a preview and obvious at the size Adobe reviewers actually inspect. Second, it was keeping tiny anchor artifacts around edges. Each one is invisible on its own. Forty files with hundreds of stray paths each is a rejection pattern reviewers recognize instantly.
The fix was boring. I stopped trusting the defaults and inspected one file end to end before batching anything. Render the traced SVG to a raster at the resolution Adobe displays it, then compare against the source. Not at 300 dpi. At preview resolution, which on my 1 vCPU box meant rendering at 60 dpi for inspection instead of 300. That one change cut my inspection render time by about 80% and caught every banding case.
Vector files have hard rules, and the reviewer does not bend them
Adobe Stock wants real vectors: clean paths, no raster embeds, no clipped raster backgrounds hiding under the art, no open paths pretending to be shapes. Vecteezy is similar but the submission metadata differs. If you are batching, you cannot hand-check 40 files for an embedded bitmap every time. You need a script that opens each SVG and fails loudly if it finds an <image> tag or a suspicious clip path.
I wrote that check in about 30 lines of standard-library Python. It has caught three raster embeds since, all from the same prompt pattern that liked to bake in a soft background texture. That pattern is now blocked at the prompt stage, which was cheaper than blocking it at the trace stage.
The metadata CSV was a separate rejection I did not see coming
While fixing the vectors I also learned that Adobe rejects a wrong CSV header before a human ever looks at the file. Adobe's upload CSV has 5 columns. Vecteezy wants 4. Dreamstime wants 15. Get one header wrong and the whole batch bounces at the door.
So make_metadata.py now writes the exact header per platform and nothing else. It walks a folder, takes the filename as a title-cased fallback, and optionally pulls real titles and keywords from a titles.json. The fallback is fine for a first pass. It is not fine for submission, and I say that in the script's own help text because I learned it the expensive way: a filename-derived title like "Blue_floral_pattern_04" reads as exactly what it is, a batch artifact.
The double-post mistake that almost got me flagged
Here is the one that still makes me wince. After fixing the trace settings, I re-ran the batch. I forgot that half the corrected files had already been uploaded the day before. Running your upload batch twice is one of the fastest ways to get an account flagged for duplicate content.
upload_tracker.py exists because of that afternoon. It is a plain ledger: pending shows what is not done yet, done marks an asset, status gives you the count. Both it and the metadata script use only the Python standard library, so there is nothing to install and nothing to break. A re-run now physically cannot double-post, because the tracker is checked before anything is queued.
The small-server constraints shaped everything
None of this ran on a workstation. The whole pipeline lives on 1 vCPU with 2 GB RAM and no GPU. That forced habits I now consider features:
- One image process at a time. Fanning out trace jobs on one core just thrashes.
-
gc.collect()after every asset, because Python holds references longer than you think. - Heavy render steps run as one file per subprocess, so memory actually returns to the OS.
- Measure peak RSS with
resource.getrusage(RUSAGE_SELF).ru_maxrss, not average usage. Peak is what kills a 2 GB box.
After those fixes, the same 40 vectors went through Adobe review with the trace issues gone. The rejections stopped being about file quality, which meant the remaining work was just titles and keywords, the part I can actually improve on purpose.
Where the tools live now
Everything in this post is a script I use weekly. They are packaged, along with the folder layout, the PDF renderer, and the product packager that zips files and renders a 1280×720 cover plus a 600×600 thumbnail, as the Pipeline Starter Kit. Four standalone Python scripts, two third-party packages total (Pillow and reportlab), validated on the same 1 vCPU / 2 GB setup. It is the tool built from this exact failure, not a course about it. If your pipeline stalls at metadata, packaging, or double-posting like mine did, it will save you the Saturday I lost.
Building digital products? Free guides & kits for digital product sellers.
Top comments (0)