A ledger of AI tools and building sites for overseas traffic.

Home / AI tools / An AI Image and Video Asset Pipeline: From Prompts to Batch Output

An AI Image and Video Asset Pipeline: From Prompts to Batch Output

Generating dozens of images makes one-by-one chat prompting unworkable. Split prompts, naming, filtering and archiving into a repeatable pipeline.

Sep 1, 2026 4 min read naket.top Updated 2026-09-07
An AI Image and Video Asset Pipeline: From Prompts to Batch Output

When you run a content site, illustrations are the most underrated part of the workload. 20 articles need 20 stylistically consistent images; generating them one conversation at a time eats a whole day.

Four stages

  1. Templated prompts: a fixed style description plus a variable part
  2. Batch generation: submit several at once instead of waiting on them one by one
  3. Automatic naming: write to disk as "purpose-index", not random filenames
  4. Filtering and compression: keep only the usable ones, crop and compress uniformly

The prompt template

The key is to write the "style paragraph" only once:

[Style] warm film texture, natural light, shallow depth of field, off-white and rust-red palette, no text
[Subject] {variable}
[Composition] side high-angle view, large areas of negative space

Only the subject gets swapped. The images that come out share one style instead of one being cool and the next warm.

Batching and naming

# Rename and compress uniformly after generation
for f in ~/Downloads/gen_*.jpg; do
  n=$(basename "$f" .jpg)
  sips -Z 1280 -s format jpeg -s formatOptions 62 "$f" --out "assets/img/${n}.jpg"
done

sips ships with macOS, so there is no need to install ImageMagick. 1280px wide at a quality around 60 is more than enough for the web.

Filtering criteria

  • Subject is clear, no garbled text
  • Matches the site's color palette
  • Does not look "AI at a glance" (extra fingers, smeared edges)

Delete the rejects immediately; don't keep them for "might use later." The value of an asset library is in quality, not quantity.

A note on cost

Generate 40 images and maybe 8 end up usable. Budget by that ratio, not by "I need 20, so I'll generate 20."

Building in 3–5x redundancy is the default setting for this kind of pipeline.

Why batching beats retouching

Polishing a single image takes 30 seconds to 2 minutes per conversation, so 40 images means one to two hours. Batch generation plus filtering afterwards gets the same 40 images in a little over ten minutes — the catch is a high discard rate. Trade quantity for time first, then trade filtering for quality.

Archiving convention

assets/img/
article/<slug>-01.jpg     article images, named by slug
series/<series>-<number>.jpg   series images, sequentially numbered
og/<slug>.jpg            share image, 1200x630

The benefit is that you don't have to open an image to find it: the filename tells you which article it belongs to.

Three things that go wrong easily

  1. Style drift within a batch: put the style paragraph at the very front of the prompt, where the weight is higher.
  2. Inconsistent image dimensions: crop everything to 16:9 or 4:3 so the page doesn't end up with ragged heights.
  3. File sizes getting out of hand: 1280px at quality 60 is typically 80–150KB; anything over 300KB should be recompressed.

In one sentence

The value of an asset pipeline isn't in how good the output looks, it's in being reproducible next time. Only once the prompt template, naming rule and compression parameters are all fixed does it count as a pipeline.

One last thing: save the prompts that failed, too. Failed samples say more about a model's boundaries than successful ones — next time you tune parameters, use them directly as a regression set.

Cost comparison

StageTime per imageTotal for 40
Prompt-only generation20–40s15–25 minutes
Generation + manual filtering+1 minute/imageabout 1 hour
Retouching one by one2–5 minutes2–3 hours

The difference is mostly in the last two steps. Batching only solves going from 0 to 1; getting from 1 to usable still takes a human pass.

AI image generationworkflowasset management

Further reading

ComfyUI docs ↗