News
Images that finally belong together
In a carousel, the first image becomes the reference for the rest: same rendering, new composition. Five slides that finally look like they come from the same place.
Ask a model for five images, with five different descriptions and the same style instruction, and you’ll get five images that have nothing to do with each other. The same palette, maybe. But studio lighting on one, raking sunlight on the next, film grain on the third. Lined up in a carousel, they tell the story of five different brands.
An anchor, then the rest
When a post goes out with its images, one image becomes the anchor: by default the first illustrated slide — usually the cover — or the one you pick. It’s generated on its own, first. Then the others go out three at a time, with the anchor’s image as their first reference and a clear instruction: match its visual style, palette, lighting, texture and mood so the images read as one series, but compose a new image for the subject at hand — and specifically, don’t copy its framing.
If the anchor fails or is stopped, the rest go without a reference rather than waiting forever. The option is called “Consistent images”, it’s on by default in the creation dialog and in an adaptation, and it states what it adds to the cost and the duration. Turned off, the label is honest: “Shared style only”.
That choice has since become a setting of the post itself. In the “Post” tab, the “Post images” section sets the “Series” — the first illustrated image, another slide you pick, or “No series” — and each image can opt out on its own: a packshot on a plain background has nothing to borrow from the others’ mood. The step-by-step is in the Guide: generate a consistent image series.
What style gives you, and what it doesn’t
Every post carries a copy of the project’s style kit: palette, fonts, corner radius, handle, logo, tone and image style. A copy, deliberately: a post keeps rendering identically even if the project’s art direction changes later.
Only part of that reaches the image model — the image style written out in plain words, and three colors from the palette: primary, accent and background. Fonts, logo and handle are of no use to it: those are layers the app puts on top. The model also gets a firm rule: no text, no letters, no numbers, no logos in the image.
A shared style kit gives you a family. A shared anchor gives you a series.
That’s exactly what the option adds. Without it, the images share colors. With it, they share a light.
Every image sized to its own area
A slide isn’t an image with text on top: it’s content and an image, two independent layers. Seven content kinds — title, title and body, list, stat, quote, call to action, visual — across six image placements: none, background, top band, bottom band, side column, frame.
The direct consequence: an image is generated at the size of its own area, not the post’s. A full-bleed 4:5 background is 1024 × 1280; a band at the top of a slide is a wide strip; a side column is a narrow portrait. Generating a square and cropping it afterwards means paying for pixels you throw away and cutting off whatever mattered.
Meanwhile, you can leave
The series runs on the server. The editor banner keeps count: “4 images underway · 3 after the reference · sized to fit their area”, with a “Stop images” button. Each thumbnail shows its state — “Image…”, “After the reference”, “Pending”.
You can close the tab. The images keep going, and land on the slides when you reopen the post, with a message: “4 images finished while you were away · Added to the post. You can undo them from the history.”
On a five-image carousel measured on Sep 17, 2026, the series without references took 25 seconds; the same one made consistent took 45 — the anchor alone first, then four edits in parallel. Consistency costs a little more, too: the reference image counts as model input, and the estimate says so before anything runs.
Measure, don’t assume
One last fix, less visible and more instructive. The server recorded the requested size as if it were the one produced. When an area isn’t a post format — a 4:5 band at the top of a slide comes out at 1504 × 864 — the model doesn’t always honor it. Nobody knew: the slide believed it was the right size, the crop quietly cut in, the “heavily cropped” warning never fired, and regenerating asked for the exact same mismatch all over again.
The size is now read from the file’s own bytes. Along the way, the chat assistant got something it had never had: the reference image and series consistency, just like batch generation. Its images came out off-style, and that was why.
