When you set up a prompt in AImitatio, there are two text fields: one for the post and one for the picture that goes with it. Most people put all their effort into the first and type a few words into the second. That's a pity, because the image prompt works differently from what you might expect, and knowing how is what makes it work.
What actually happens to your image prompt
On each scheduled run, the scheduler picks one of your prompts, writes the post, and then sends the image prompt, exactly as you wrote it, to FLUX.1-schnell running on our own GPU. Two things are not sent along:
- Your Context. The briefing about who you are and how you sound only goes to the text model.
- The post that was just written. The image model doesn't see today's text.
So the image prompt has to stand on its own. If your Context says "we are a small bakery with a rustic look" but your image prompt just says "bread", the image model only ever sees "bread".
Rule 1: Put your look into every image prompt
Decide on a short "style line" once and add it to the end of every image prompt in your app. For example:
"…, natural window light, warm muted colours, wooden surfaces, photographed from slightly above, no text."
Same style line, different subjects: that's what makes a feed look like one account instead of a random collection of pictures.
Rule 2: Describe a scene, not a message
Image models are good at things, places, light and composition, and they don't know your brand. A prompt like "our commitment to quality" gives the model nothing to draw. Describe what someone would actually see:
- Subject: "a sourdough loaf, cut open"
- Setting: "on a flour-dusted wooden board"
- Light and mood: "early morning light from the side"
- Framing: "close-up, centred"
Leave out logos, specific products and real people. The model doesn't know what yours look like, and an invented version is worse than none. Ask for no text in the image too: the words belong in the post.
Rule 3: Write for every post the prompt can produce
The image prompt is fixed per prompt, but the text comes out new on every run. An image prompt that only fits one particular post will clash with the next one. Keep the picture one level more general than the text: if the text prompt is "share one tip about keeping bread fresh", an image of a loaf wrapped in a linen cloth fits almost any tip it produces.
Rule 4: Think square
Scheduled images are generated at 1024 × 1024 pixels. Prompts that describe one centred subject usually work better in a square than wide panoramas or tall scenes.
Test before you schedule
You don't have to wait for the next scheduled run to see what an image prompt produces. Open Generate, paste the image prompt with your style line, and run it. It uses the same FLUX model and the same 1024 × 1024 default size as the scheduler. It's free with an account, one job at a time. Run it two or three times to see how much the results vary. If they don't fit, change the prompt there, then copy the final version into your prompt in the app. You can edit an existing prompt's image text at any time from your prompt list.
Good to know
- If our GPU can't produce an image on a run, an external image service is tried as a fallback. If that fails too, an image generated earlier for the same image prompt may be reused, and otherwise the post goes out without a picture. The log e-mail after each run tells you which happened.
- On Bluesky, about one run in ten tries to turn the generated image into a 3-second video with Wan2.2, using the same image prompt. A gentle, natural motion (steam, light, leaves) suits that better than a busy scene. Videos made this way are also published to our public video gallery.
- Leaving the image prompt empty is fine. Not every post needs a picture.
Checklist
- Every image prompt describes a visible scene: subject, setting, light, framing.
- Each ends with the same style line and "no text".
- No logos, product look-alikes or real people.
- Each fits any post its text prompt can produce.
- Tested in Generate before it went into the schedule.
Want to see how your image prompts come out? Create a free account, try them in Generate, and add them to your first prompts.
