Template or generated image? It depends what has to be written

There is a simple rule for choosing between generating an image with AI and rendering a template, and it fits in one sentence:

If the piece has a word that has to be spelled right, it is a template. If the piece is a scene, it is an image model.

The rest of this piece explains why that rule works, and what changes when you stop treating the two as competitors.

What image models do well

A lot. Modern image models produce scene, light, texture, depth and composition at a level that was impossible not long ago. Bread coming out of the oven with steam and morning light; an office desk with the mug at the right angle; a fabric texture for a background. That is expensive to photograph and cheap to generate.

For those pieces, templates do not compete. You are not going to draw steam in CSS.

What image models do badly, and probably will keep doing badly

Text. It has improved a lot, but the problem is structural: the model draws something with the shape of text, it does not write text. On a short common word it usually gets it right. On a word from your business — your brand name, a price with decimals, a date, a product name — the error rate climbs. And the error is the kind nobody accepts publishing: a swapped letter in your own company name.

Your exact colour. You ask for your indigo and get a nearby purple. On one piece, nobody notices. Across twenty pieces in the same feed, the whole feed looks like another brand — because consistency is what makes a visual identity work, and “nearby” is not consistent.

Your font. There is no way. The model draws letters that resemble a typeface; it does not use yours.

Position. “Price bottom right, logo at the top, legal notice in the footer” is a layout instruction, and image models do not lay out — they compose. The result lands somewhere near the request, and “near” moves with every generation.

What a template is, technically

In EverFeed, a template is three things: HTML, CSS and a JSON Schema. The HTML is the structure of the piece, the CSS is the design, and the schema describes the fields that piece needs — title, price, CTA, background image. The schema becomes a form on its own: whoever uses the template just fills in the fields.

At generation time, the engine assembles the final HTML and opens it in a headless Chromium, which photographs the result as PNG, JPG or MP4. It is the same engine that draws a web page, so everything the web knows how to do — typography, grid, shadow, gradient, mask, animation — is available.

And here is the part that changes the game: colour and font are never hard-coded into the template. They arrive as CSS variables injected from the company brand kit. The template says “use the primary colour”; who says what the primary colour is, is the client record.

The consequence: one template, every brand

Because colours and fonts come from outside, the same template file renders with each company identity. An offer template designed once serves the bakery, the clinic and the parts shop — each with its own indigo, its own green, its own heading font, its own logo.

For anyone running a single brand, that means the gallery of ready-made models already shows up looking like you: the thumbnail displays the template in your colours, not the sample ones. For anyone running five brands, it means improving one template improves all five at once.

That is impossible with an image model, and not for lack of effort: the brand is not a model parameter, it is a description in a prompt. Descriptions approximate.

Cost, which nobody brings into this discussion

Rendering a template costs the time it takes a headless browser to open a page. In practice, zero per generation. Generating an image costs per piece, and the wrong attempt costs the same as the right one.

That inverts the economics of trying. Adjusting a template is free and deterministic: change the text, render again, get the same thing with new text. Adjusting an image prompt is paid and probabilistic: change the prompt, generate again, get something else — sometimes better, sometimes not, and the good version you had before may never come back.

That is why, in the arithmetic of a month of content, the image line halves when you move carousel cards to templates. It is not a saving of cents: it is taking half your pieces out of a roulette they should never have entered.

The decision table

The last row is the most important

Both in the same piece is the most common case, and the worst handled in tools in this category. The model generates the scene; the template takes that image as a background and writes on top of it, with your font, in your colour, in the exact spot. You get what each technology does best and pay for neither weakness.

In EverFeed that is not a configuration trick: it is how the pipeline was designed. The visual step can come from an image model, from a template, or from a template that uses the model output. And because each step is a separate row, you can redo just the background while keeping the layout — or lock the layout and test five backgrounds.

When to break the rule

Twice, in our experience.

When the text is decorative. If the word in the image is ambience — a sign in the background, an out-of-focus shopfront — let the model draw it. Nobody is going to read it.

When the piece is a test. To find out whether a concept works, generating fast is worth more than generating right. Just do not publish the test: turn the approved concept into a template and publish the template.

Outside those two cases, the rule from the first sentence holds. If it has to be spelled right, it is a template.