Image generation pipeline: from the task to the file on the site
Seven steps: the task, the prompt, a series, selection by criteria, editing by mask, upscaling and finishing — one master for every format and text in the layout, with live examples.
In short
A good generated image is not one lucky prompt but a pipeline of seven steps. First the task: where the image will be shown, in which formats, where the text goes. Then a prompt by the formula and a series of 8–16 variants, in which one thing changes at a time. The best one is chosen by criteria set in advance — the product is right, hands and edges are intact, there is no letter soup — and not by the first impression. The chosen image is edited by mask instead of being regenerated, upscaled to twice the largest display size and cropped from one wide master into every format. Text and logos are added in the layout, colour is matched to the series, and the files go to the site as AVIF or WebP with a description. The prompt, the seed and the edits are saved, so the series can be continued a month later.
Seven steps from the task to the file
Press a step: what goes in, what comes out and what to check before moving on.
01Task
- In
- where it will be shown and why
- Out
- format, size, room for text, 3–5 references
- Check
- the proportions and the room for a headline are known before the first prompt
02Prompt
- In
- the task and references
- Out
- a prompt by the formula
- Check
- one scene; light, camera and format are named
03Series
- In
- the prompt
- Out
- 8–16 variants
- Check
- one word changes at a time; where possible, the seed is fixed
04Selection
- In
- the series
- Out
- 1–2 candidates
- Check
- criteria in advance: the product is right, hands and edges are intact, no letter soup
05Editing
- In
- a candidate
- Out
- a fixed image
- Check
- inpainting by mask: a detail, the background, the frame extended for a format
06Upscale
- In
- the edited image
- Out
- a file of the right size
- Check
- twice the largest display size, no “plastic” skin
07Finish
- In
- the final image
- Out
- files for the site and social media
- Check
- colour matched to the series, text and logo in the layout, AVIF or WebP, alt
Most of the time goes not into the prompt but into selection and editing. That is why the first step matters most: when the format, the room for text and the criteria are known in advance, the series is generated for the task, and the selection takes minutes instead of an evening.
The task for an image: six questions
The answers fit in a few lines and save a dozen series.
| Question | Example answer |
|---|---|
| Where will it be shown? | a banner on the home page and a 4:5 post |
| What is the main thing in the frame? | the whole armchair, the fabric texture visible |
| Where does the text go? | the left third of the frame is empty |
| What light and mood? | soft daylight, a light calm interior |
| What is not allowed? | people, other brands’ logos, inscriptions |
| What are the references? | three photos from the brand’s catalogue |
A series and selection by criteria
Eight variants of one prompt. Six are rejected for a reason that can be named, one is chosen, one is kept in reserve.
If the generator allows it, the seed — the starting number of the generation — is fixed, and the series changes one word at a time: then it is clear what exactly changed the image. If nothing fits after two series, the prompt is fixed, not the luck: the same reasons for rejection repeat and point to the missing words.
Edit instead of regenerating
A good composition with one broken detail is fixed by mask — the rest of the image stays as it is.
A light grey wool armchair in a bright Scandinavian living room, soft daylight from a large window on the left, eye-level shot, 35mm lens, calm neutral palette of sage green and off-white, wooden floor, a small plant on the right, photorealistic interior photography, the armchair placed slightly right of centre, aspect ratio 3:2Inpainting replaces a selected area: an extra leg, a strange hand, a stain on the wall. Outpainting extends the frame: a square becomes a wide banner, and the room for the headline appears where it was missing. Both are built into many generators and into Photoshop as generative fill and generative expand.
A new generation is needed when the composition, the light or the angle are wrong: no mask will turn the light from the right into light from the left. Then the reason goes into the prompt, and the series starts again.
Upscaling comes after editing: the image is enlarged two to four times so that it stays sharp on the largest screen. A “creative” upscale draws new details — faces, fabric and text are checked afterwards.
One master — every format
Switch the formats. A wide master with the subject in the middle third gives a post, stories, a cover and a link preview without new generations.
That is why the format is asked about at the first step: a vertical story cannot be cut from a frame where the armchair stands at the edge. If the master is too narrow for a banner, the frame is extended at the editing step.
Text and logos — in the layout
Generators get better at short words, but the brand’s name, prices and logos are added on top in the layout.
In the layout the text is set in the brand’s font, does not melt, can be fixed in a minute and translated into another language. In the prompt it is enough to leave a clean surface: “a plain kraft bag without any print”.
The pipeline on file: 2 examples
A record for the prompt library that lets you repeat the image, and the markup with which it goes to the site.
A record in the library
One record per published image: the prompt, the seed, the edits and where it is used.
# Enough to repeat the image or continue the series
id: armchair-living-room-01
used: home page banner, 4:5 post
generator: the name and version of the generator
prompt: >
a light grey wool armchair in a bright Scandinavian living room,
soft daylight from a large window on the left, eye-level shot,
35mm lens, calm neutral palette, empty space on the left third
aspect: "3:2"
seed: 48213
edits:
- inpaint: removed the extra back leg
- outpaint: extended the left side for the banner
upscale: 2x
final: /img/armchair-living-room-01.avifOn the site
AVIF and WebP with a JPEG fallback, three sizes for different screens, the size of the image and a description.
<picture>
<source type="image/avif" sizes="100vw"
srcset="/img/armchair-960.avif 960w, /img/armchair-1920.avif 1920w, /img/armchair-3840.avif 3840w">
<source type="image/webp" sizes="100vw"
srcset="/img/armchair-960.webp 960w, /img/armchair-1920.webp 1920w, /img/armchair-3840.webp 3840w">
<img src="/img/armchair-1920.jpg" width="1920" height="1280" fetchpriority="high"
alt="A light grey wool armchair by a large window in a bright living room">
</picture>7 rules of the pipeline
-
01
The task before the prompt
Formats, room for text and what must not be in the frame.
-
02
A series, not one shot
8–16 variants, one change at a time.
-
03
Criteria in advance
Every rejection has a reason that can be named.
-
04
Edit by mask
Regenerate only when the composition or the light are wrong.
-
05
A wide master
The subject in the middle third — every format is cut from one image.
-
06
Text in the layout
Names, prices and logos are not generated.
-
07
A record for every image
The prompt, the seed and the edits — to repeat the series.
Common mistakes in the pipeline
-
Generating before the task
A beautiful square that does not fit the banner.
-
Choosing by the first impression
The extra finger is noticed by the client, not by you.
-
Regenerating because of one detail
A good composition is lost for the sake of one hand.
-
Text drawn by the generator
Letters melt, and a typo cannot be fixed.
-
An image without a size
A 6 MB PNG on the home page and a jumping layout.
-
Nothing saved
A month later the series cannot be continued.
Questions about the image pipeline
How many variants should be generated?
8–16 per idea. If nothing fits after two series, the prompt is fixed, not repeated.
Edit or regenerate?
Edit when one detail is wrong; regenerate when the composition, the light or the angle are wrong.
What resolution is needed?
About twice the largest display size: a banner shown 1920 px wide needs a source of about 3840 px.
Can text be generated right in the image?
Short words sometimes work, but brand names, prices and logos go in the layout: the exact font, easy to fix, easy to translate.
How do I keep a series consistent?
The same words for light and palette, a fixed seed or a style reference where the generator supports it, and a colour pass at the end.
Which format is best for the site?
AVIF or WebP with a JPEG fallback, several sizes through srcset, width and height set and a description in alt.