Featured images and diagrams

Generating images, overlaying text, using reference images (logos, characters), and regenerating failed images

Featured image

Sorabun automatically generates an image matching the article's title. Under "Settings > Generation defaults > Featured image," you can set:

• Generate automatically by default (eyecatch_enabled)
• Overlay the title text (eyecatch_text): images with text overlaid tend to get a higher click rate on social media and in listings
• Style (eyecatch_style): flat, photo-realistic, and so on

Sorabun composes the on-image text before handing it off

Passing the title straight to the image generator causes problems like these.

• In 「LINE公式アカウント料金プラン改定｜26年10月の変更点と節約術」, the "｜" gets drawn straight into the picture. It's a meaningful character in search results, but it has no meaning inside an image
• The image generator decides where to wrap the line itself, and since it doesn't understand Japanese word boundaries, it can break a word in the middle (「…料金プラン改」／「定｜26年10月の…」)

So Sorabun builds the text that goes on the image from the title before handing it over.

• Separators such as ｜ | ／ 【】 become the line break. The symbol itself isn't included
• When there's no separator, Sorabun picks a break point in this order of preference: punctuation, then a particle boundary, then a change in character type, choosing whichever candidate falls closest to the middle
• It avoids starting a line with punctuation, a small kana (っ・ゃ・ゅ・ょ), or a particle (が・を・に・へ・の)
• Even a long title is capped at 3 lines. Anything beyond that isn't readable at thumbnail size

Illustrative images

Separately from diagrams, you can add a picture that brings the scene to mind, placed right in the body text.

For an article that walks through a procedure, the two split the work like this.

In a 5,000 character article with only two diagrams, the rest of it is a wall of text. Reading-style articles, personal stories, and opinion pieces can end up with no diagrams at all (the AI decides a diagram wouldn't be effective there). That's exactly where this feature helps.

Settings

Found under "Settings > Generation defaults > Number of illustrative images" (photo_count).

• Don't include (default)
• AI decides: adds an image only where the scene is hard to picture from the text alone (up to 3 images). If the text already carries the scene everywhere, it can come out to 0 images
• 1 to 5: sets a fixed number of images

Art style (photo_style) offers the same 4 choices as the featured image (flat illustration, photo-realistic, minimal, gradient). Which AI generates them is set at "Settings > Models > Image generation > Illustrative image" (photo_provider).

Design decisions

• No text is drawn into the image. Japanese characters are prone to coming out garbled, and garbled characters in a photo-style picture stand out far more than a garbled diagram does. Even in a scene with a screen or a sign in it, that surface is left blank or abstract
• Sections that already have a diagram are skipped. Two pictures stacked under the same heading leaves the reader unsure which one to look at
• Not added for abstract discussion, definitions, or lists of numbers. A picture wouldn't add any meaning there
• If an image can't be created, that one image is skipped and the article is finished anyway. An article is complete with or without illustrative images

Diagrams

Sorabun generates and inserts flowcharts, step diagrams, and comparison diagrams based on the body content.

• Number of diagrams (diagram_count): 0 to 5, or "Let AI decide"
• Aspect ratio (diagram_aspect): 16:9 by default

With "Let AI decide," diagrams are only inserted where the explanation is genuinely complex, so you don't end up with meaningless diagrams.

Image inspection (Pro)

Sorabun shows the generated image to an AI to check for obvious problems (three-fingered hands, extra limbs, garbled text) and regenerates it if it finds any (eyecatch_check, defaults to 1 retry).

If the problem isn't fixed after a retry, it's left as a warning on the job. Treat that as a cue to replace the image yourself.

Note that Sorabun is already instructed to avoid close-ups of hands and fingers, which are prone to these errors. It prioritizes objects, scenery, and symbols over people, and when people do appear, keeps them small and shown from behind.

Regenerating a failed image later

If only the image generation fails for an article, Sorabun stops it as a draft instead of publishing it. Otherwise the article would go out to readers and search engines with a missing image. The reason it stopped is left as a warning in "Jobs / History."

A stopped article can be regenerated from the WordPress posts list. Hover over the row and a "Regenerate Image" link appears. Click it and Sorabun does the rest. The same link appears for pages too.

• Featured image: regenerated using the style, reference images, and text overlay settings you already have configured, then set on the article automatically
• Diagrams (Pro): rereads the current headings in the body and chooses new spots for the diagrams. Even if you've edited the headings since the article was first generated, the diagrams still land in the right places

The article stays in draft even after it's regenerated. Check the images, then publish it yourself.

"Regenerate Image" also appears for articles that simply have no featured image, such as ones created before you added an API key. Add the key later, then regenerate them all at once.

Regenerating an image you don't like

Even when nothing has failed, you can end up not liking the finished image. For articles Sorabun created, the row also offers an entry point for that case.

These are kept separate by type. If they were combined into one button, regenerating just the featured image would also regenerate the diagrams, adding cost you didn't need.

Regenerating just one image

When you don't like just one diagram or illustrative image, you can replace only that one.

On the "Sorabun > Images" list, click "＋ Regenerate one at a time" on the article's row, and the images in that article are listed. Click "Regenerate this one" on the one you want to fix.

• Only that one image changes to a new picture. The other images and the body stay the same
• The original content stays in WordPress's revision history, so you can revert at any time
• It's regenerated with the same content, so for a diagram, the flow and the items stay exactly the same, and only the artwork changes

Regenerating in bulk (the Regenerate Images screen)

If you generate 10 articles and 3 of them stop because of images, clicking each one individually isn't realistic. "Sorabun > Jobs / History > Regenerate Images" collects the articles that are missing images.

1. Open "Sorabun > Jobs / History" and choose "Regenerate Images" from the tabs at the top
2. Check the boxes for the articles in the list (check the header box to select all)
3. Click one of the buttons below
   - Create missing images
   - Regenerate featured image
   - Regenerate diagrams (Pro)

For each article, the list shows whether it has a featured image (and what that image is), how many diagrams it has, and which type of generation failed. Use "Articles missing images / All articles" at the top to switch views. Only articles Sorabun created appear in the list.

Articles that don't qualify (nothing to regenerate, or no permission) are silently skipped, and the screen shows how many were started and how many were skipped.

Using reference images (logos, characters)

You can have your own logo or character appear in generated images.

1. Upload the image to the media library
2. Specify it in the reference image field under "Settings > Generation defaults" (eyecatch_ref_ids / diagram_ref_ids)
3. Choose how it's used
   - Always use it: includes it every time
   - Leave it to AI: only uses it when it fits the content

Sample images: match the look of your older articles

If you have featured images or diagrams from before you started using Sorabun, you can register them as "samples." New images then lean toward their colors, technique, line weight, shading, and typographic feel, so when old and new articles sit side by side in the posts list, the ones Sorabun made don't clash with the rest.

1. Put the images you want to use as samples (older featured images, older diagrams) in the media library. An image already attached to an article can be picked as it is
2. In the sample image row under "Settings > Generation defaults," choose them separately for featured images and for diagrams (eyecatch_sample_ids / diagram_sample_ids; up to 2 of each)
3. This applies to images generated from then on. Images already generated don't change

This differs from reference images (logos, characters) in whether the subject actually appears in the image.

Samples aren't handed to the image model as they are. Passing the image itself would copy over its content along with its look, producing "an image that's a dead ringer for the sample." Instead, the AI (Gemini) looks at the sample first and turns its colors, technique, line weight, lighting, and typographic feel into a short English sentence, and that sentence is what goes to the image model. Since the sentence never says what's depicted, there's nothing for it to copy. This extraction happens only once, when you change the sample, and the extracted sentence appears next to the sample field on the settings screen as "Extracted look." Extraction requires a Gemini API key (not needed on the Cloud edition).

For character reference images, Sorabun instructs the AI to draw the character facing forward with the face visible, in the same pose, expression, and colors as the reference image. The rule that avoids showing people (keeping them small and turned away) doesn't apply to characters. During the pre-publish check, if a character comes out facing away or as a silhouette, Sorabun regenerates it.

Choosing an image generation AI

Under "Settings > Model," you can choose Gemini or OpenAI (GPT Image) separately for featured images and diagrams.

OpenAI's models (GPT Image 2.5)

You choose this with openai_image_model. The default is gpt-image-2.5-flare.

Sorabun generates the featured image and diagrams together for each article, so the default favors speed. If you'd rather take your time over each image, switch to sunburst. The field accepts free text, so when a new model comes out, you can just type its name in.

With 2.5, openai_image_quality gained two more levels: xhigh (highest) and max (maximum). The higher you go, the more it costs.