AI social posts from your own photos (without letting it invent the picture)

Almost every AI content tool sells the same thing: describe the topic and it gives you copy and images. For a blog or an opinion account, fine. For a bakery, a garage or an estate agency, the generated image is precisely the part that does not work — because the photo does not accompany the message: the photo is the message.

This article is about the opposite: using AI for what it is actually good at — writing, adapting per network, ordering the week — and supplying the photos yourself.

Why can people tell a photo was generated by an AI?

It is usually not an obvious glitch. Current generators stopped drawing six-fingered hands a while ago, and by now hardly anyone spots a generated image looking at it in isolation. What gives it away is far simpler:

It is not your place and it is not your product.

A customer who has been to your shop can see that is not your counter. A customer who has bought from you can see that is not your packaging. And someone who has never been sees a generic stock image, which is exactly what a brand that does not exist looks like. On local social media — where a good share of your readers have walked past your door — the invented photo fools nobody: it reads as "a machine wrote this", and that taints the copy too, which may well have been fine.

Two more practical consequences:

• What you publish is what you can reuse. A real photo of your product works for the post, for the product sheet, for the catalogue and for answering a customer in a message. A generated image only works for that one post.
• Platforms label generated content. Meta adds an "AI info" label to content it detects through the invisible markers and metadata (C2PA, IPTC) that generators leave behind, or when whoever uploads it says so (Meta Transparency Center, checked in August 2026). It is not a penalty, but it is a public label next to your brand that you did not choose to put there.

The conclusion is not "AI is no good for social media": it is that it pays to be careful about which half of the job you hand over. Writing seven different texts, adapting them per network and spreading them across the week is repetitive work and AI does it well. Inventing your shopfront, no.

Can the AI write the copy while I supply the photo?

Yes. In the PlanVortex planner it is called generating the plan from your photos (from_images in the API):

1. You upload the photos you want to publish, each with one line saying what is in it.
2. The model reads all of them at once, in a single pass, together with your topic and your brand context.
3. It returns the week: which photo goes on which day, the copy of each post adapted to the network it lands on, and the time.
4. You review, fix whatever you want and validate. No image is generated.

That the pass is one and not one per photo is not an implementation detail: it is what lets the week have a thread. The model sees the twelve photos together, so it can notice that number 3 is the "before" and number 7 the "after", and order them so the week tells something. Photo by photo you would get seven independent posts that happen to be consecutive.

How do you describe a photo so the copy works?

This is the part that decides whether the result is publishable, and where you win or lose the time you saved. The description is mandatory on purpose: without it the model writes about what it believes it sees, and with a product photo it gets the product wrong about half the time.

Four rules that make a visible difference:

1. Say what it is, not what it looks like. The model can already see the photo. What it cannot know is what it is looking at.

Instead of	Write
"Nice photo of pastries"	"Butter croissant, baked at dawn, gone before eleven"
"Our garage"	"Timing belt replacement on a Golf VII, two hours of work"
"Flat in the centre"	"3-bedroom flat on Calle Mayor, refurbished in 2025, third floor with no lift"

2. Proper nouns. The name of the dish, the model, the street, the reference. It is literally the only thing the model cannot infer from the image, and it is what makes the copy sound like yours instead of like a template.

3. What the picture does not show. The price, that it is seasonal, that it is only there on Saturdays, who it is for, how long it takes. A photo says none of that, and it is half of what makes a post useful.

4. One or two sentences, not a paragraph. Up to 500 characters per photo are kept; past that it is trimmed. And more is not needed: the description is context, not the copy of the post.

A trick that works: write the description the way you would say it over the phone to someone who is going to write the post for you and has not seen the photo. That is exactly what is happening.

How many posts come out of N photos?

One per photo at most, and always bounded by the week. Two limits at once:

• The source: each photo is published only once, so 12 photos are 12 posts as a ceiling. The maximum per plan is 20 photos.
• The week: the slots are the days you picked multiplied by the accounts in the plan. There is at most one publication per day and account.

The smaller of the two wins. With 12 photos and one account publishing Monday to Friday there are 5 slots: 5 get published and the other 7 stay out. The plan is weekly and the source does not extend it, so the leftovers are not piled up at the end: they stay for next week, which is where you would want them anyway.

What matters is that you know this before generating, not after: the wizard warns you while you pick the photos and repeats it at the review step, and the generated plan carries that notice inside it (code 2117 if you integrate against the API). A plan that charges you and then publishes half of what it showed you is worse than a plan that warns.

And the order matters: each photo's position is the narrative. The plan stores which photo each publication used, so a retry does not reshuffle them — you do not end up with the dessert photo carrying the starter's copy.

What does a plan from your own photos cost?

Much less, and not by a little. In PlanVortex AI usage is measured in credits included in your plan, and these are the two bills for the same week of 7 posts with a picture, on one account:

Item	Generated images	Your photos
Planner pass	15	20 + 2 per photo = 34
Copy (7 × 2)	14	14
Images (7 × 70)	490	0
Total	519 credits	48 credits

The whole difference is one row. Generating an image costs about 35 times what writing the text beside it costs, so in a plan with a picture on every post the images are around 94% of the spend. Removing them leaves the plan at something you barely notice in the monthly allowance.

The planner pass does go up a little — from 15 to 34 — because analysing the photos costs: each image adds around 1,100 input tokens, a fixed budget that does not depend on whether you upload a 12 MP phone photo or a 1,000-pixel one. It costs 19 credits more and saves 490.

What if you do not have a photo of everything?

There are two ways out before going back to generated images:

• From your catalogue. If you sell products and have a catalogue connected to the account, the plan can be generated straight from it: it takes the name, the description, the price and the real picture of each product you pick, and builds the week around that. It does not generate images either. It is the option no generic AI tool can offer, because it needs the connection to the network's catalogue.
• With no picture. A post can go without an image on almost every network. The exception is Instagram, which requires a file: there, no photo means no publication.

And for whatever you have neither photographed nor in a catalogue, the usual template is still there and does generate the images. The decision is not "AI or no AI", it is what each week comes from.

How to do it, from the panel or the API

In the panel it is the planner wizard: you choose what the plan is generated from, upload the photos with their descriptions, pick accounts and days, and review before validating. What it costs is shown before anything is generated.

If you integrate against the API, it is the same create-a-plan endpoint with two more fields:

POST /v1.0.0/clients/{id_client}/organizations/{id_organization}/ai_plans

{
  "prompt": "Artisan bakery, warm tone",
  "accounts": ["66d04a6a427f4c43b9d97f54"],
  "template": "from_images",
  "source": {
    "images": [
      { "id_upload": "66d0…", "description": "Butter croissant straight out of the oven" },
      { "id_upload": "66d1…", "description": "The bakery at five in the morning, kneading" }
    ]
  }
}

The available templates, what each one accepts, its fields and what its orchestration costs are all read from GET /planner_templates: they are published precisely so you do not have to copy them into your code and so they cannot go stale on you. You can see the detail in the documentation or what each network does. And if you would rather ask for the week out loud than fill in a form, the same thing is handled from an MCP server in your AI assistant.