Pinterest reference search inside the agent

There is a specific kind of prompt that never converges. You write "soft natural light, muted earthy palette, slightly desaturated, film-like grain, cozy but not twee," you get back something that technically satisfies every clause and looks nothing like the thing in your head, and you add three more adjectives. Five generations later you have a paragraph that reads like a mood board caption and an image that reads like a stock photo.

The problem isn't your vocabulary. It's that a visual style has more information in it than a sentence can carry, and every round of adjective-stacking is you trying to compress a picture into words so a model can decompress it back into a picture. Two lossy conversions in a row. search_pinterest skips both of them.

What the search actually returns

The tool takes a query and an optional limit, searches Pinterest, and returns image URLs. That's the whole surface. It's a lookup step rather than a generation, and what comes back is a set of real addressable images rather than a rendered board you have to screenshot.

That distinction matters more than it sounds. The output being URLs is what makes the next step possible: those images can go straight into a generation as references, or into a slideshow, without a download-and-reupload detour. Searching Pinterest for inspiration is the capability page, and it's worth reading the three steps there as a loop rather than a one-shot: query, look, requery. You are not trying to find the image. You are trying to find the two or three that agree with each other about a treatment.

The habit: brief with the picture, not about it

Once you have reference URLs in the conversation, the briefing shape changes. Instead of describing a look, you attach it and describe only the delta.

Compare these two:

"A skincare bottle on a bathroom counter, soft natural window light, muted palette, slightly desaturated, shallow depth of field, morning feeling, editorial but warm."

"[reference image] Match this lighting and colour treatment. Subject is my skincare bottle on a bathroom counter."

The second is shorter, more specific, and far more repeatable. The words in it are doing the job words are good at — naming the subject and the placement — and the picture is doing the job pictures are good at, which is everything else.

There are three concrete routes from a reference URL to a generation, and they are not the same thing:

1. Image-edit models. A substantial subset of the catalog's image models are edit models that require at least one input image and will reject a generation submitted without one. When you pick one of these, the reference isn't optional garnish, it's a required input.
2. Image-to-image. generate_image_from_image takes a prompt plus a source image, with optional additional reference images and a mask. This is the path when you want the reference to be the thing being transformed rather than the thing being imitated.
3. Reference-to-video. Video models built around reference inputs use a still to fix identity, product or set before any motion happens. The best reference-to-video models covers which ones are actually built for it.

The one thing to be deliberate about is what you're asking the reference to control. A reference image can carry a colour treatment, a composition, a lighting setup, or an identity, and if you don't say which one, the model will make its own choice and often pick the most literal one. "Match the lighting" and "match the composition" produce very different results from the same picture.

Query craft: search the treatment, not the subject

The instinct is to search for your own subject. If you're making a skincare ad you search "skincare product photography," and you get back a thousand images of skincare products shot every possible way, which tells you nothing about which way you want.

Search for the treatment instead. "Overcast window light still life." "Warm film grain interior." "High-key white product on white." These queries return images across many subjects that share the one property you're actually trying to specify, which makes the shared property obvious and makes it much easier to spot the two images that genuinely agree.

A few query patterns that tend to work:

Instead of	Search	Because
"coffee brand ad"	"morning kitchen natural light"	Isolates the lighting from the category
"skincare photography"	"wet surface macro reflection"	Isolates the technique
"gym content"	"high contrast harsh shadow portrait"	Isolates the mood
"product on table"	"overhead flat lay muted palette"	Isolates the composition

The tool description names the categories it's genuinely good at: product photography, aesthetic boards, design references, and trending visual styles. Those are all treatment-shaped, not subject-shaped, which is a hint about how to phrase the query.

If you know what format you want but not what it looks like, the prompt format library is a faster starting point than a search, since the entries there are already written as generation prompts. POV hands holding a product is a good example of a format that's easier to point at than to describe.

The chaining trap

This is the one landmine in the workflow, and it will silently waste a step if you don't know about it.

Versely's slideshow tools are end-to-end by design. create_slideshow plans its own per-slide search queries, curates a matching image for each slide, and applies overlay text, all in one call. Its own guidance is explicit: do not chain it with search_pinterest or with image-generation tools. The caption-in-image variant carries the same warning, plus a warning against stacking a text overlay on top of text the image model already rendered.

So the rule is: Pinterest search is for briefing a generation, not for feeding a slideshow. If you're building a slideshow, describe the theme and let the slideshow tool do its own curation. If you're building a single generation or a video, search first and brief from the results. Mixing the two produces double work at best and a doubled text layer at worst.

Making a good reference durable

The worst outcome of a successful reference search is that you find the perfect image, use it once, and lose it. Three things fix that, and they're separate on purpose:

• Workflow assets are for the recurring cast. If a character, product or location has to appear consistently across scenes, it belongs in an asset rather than pasted into each prompt. Setting up reusable characters and products walks the upload path.
• The brand kit is for brand-level constants: colours, fonts, logo, tone. Setting your brand kit means those stop being something you restate in every brief.
• Long-term memory is for durable preferences that aren't assets. "Always make my videos 9:16," "keep captions minimal." Teaching the agent a lasting preference is the right home for those, and explicitly not the right home for brand assets — those go in the kit.

Get this split right and the reference-first habit compounds. The first project costs you a search and a decision. The tenth one inherits both.

FAQ

Where does the spend land in a reference-first workflow?

On the generations, not on the looking. A search is a lookup step that returns URLs; the credits go to whatever image or video you produce from the reference, priced on whichever model runs it. That is the argument for doing the reference work before you start generating rather than after, because a round of searching replaces a round of adjective-stacking and only one of those two dispatches a model.

Can I use a reference image to copy someone's style exactly?

You can point a model at a reference and ask it to match a treatment, and that's a normal part of art direction. Where it stops being fine is copying a specific creator's distinctive work rather than a general visual approach, which is a rights question rather than a technical one. The practical version: use references to specify lighting, palette and composition, not to reproduce a particular image.

Should I use one reference or several?

Several, when they agree. Two or three images that share the property you care about make the target much less ambiguous than one image that carries every property at once. The failure mode with a single reference is that the model picks the wrong thing to copy from it, and the failure mode with too many is that they contradict each other and average into mush.

Why can't I feed Pinterest results into a slideshow?

Because the slideshow tools already do their own curation end to end, including planning the per-slide queries. Handing them pre-searched images duplicates a step they're going to perform anyway. If you want specific images in a slideshow, the supported path is the one built for your own uploads, not a search chained into the automatic curator.