FiftyFlowers · Marketing

Consistent product photos, on demand.

Generative product-photo editing with reusable looks: new backgrounds, exact brand colors and transparent cutouts, with the cost of every run shown before it is spent.

PROJECTImage Studio

MY ROLEAI Solutions Engineer · FiftyFlowers

STAGEIn production · released August 2026

An Image Studio result: the original photo of red roses in a vase beside the same roses on a flat grey background

Image Studio: the original beside the result. A local run of the product’s code; the result shown is the sample that ships with the look, and no model was called.

The context

A problem worth solving.

A large product catalog needed consistent backgrounds and styles. Editing by hand was slow, and a credit-based third-party tool cost about 15 cents an image.

My contribution

What I brought to the work.

Ran the April pilot and its cost case, then built Image Studio and Preset Creation into the company’s AI platform, with usage tracking and team limits.

The product

What it makes possible.

Five parts follow: Image Studio, Preset Creation, costs and limits, how results are checked, and the April 2026 pilot.

A pilot that made the case

Five catalog products on five backgrounds gave 25 image-to-image results: the simplest product kept its shape on every background, while most tight flower close-ups were recomposed. The pilot estimated about 4 cents and 5 seconds an image on the model it tested.

Pick a look

Upload a batch and apply a saved look: background color, vase, lighting, backdrop texture and one of 14 aspect ratios.

Exact color, real cutouts

An exact-color option rewrites background pixels to the brand color after generation; the cutout look produces a transparent PNG.

Looks the team can build

Preset Creation derives a reusable look from example images and a test photo, generates three samples, and keeps up to ten versions.

Costs in view

Every run shows photos times the price the administrators have set and asks for confirmation above 25 cents; administrators see image spend over time and set per-person and team limits.

At a glance / simplified product view

01Product photo
02Saved look and AI model
03Styled image or cutout

Image Studio · in the company AI workspace

Upload a photo, pick a look, keep the same flowers.

In production · released August 2026

A team member uploads product photos, picks a saved look and downloads the same flowers on a consistent background. These pictures are the workspace’s own code running on my machine with invented accounts and a stand-in in place of the image model. No image model was called: a result shown is either a sample that ships with the app or a tile marked as a placeholder, and each caption says which.

The Image Studio first screen with a photo drop area and a row of six looks — **Start with a photo and a look**The first screen: a drop area for product photos and the row of saved looks, five built in and one made by the team in this run. Local run, invented accounts; no image model was called.

A look’s options with colour, backdrop, lighting and composition choices above a Generate button and a cost line — **Change a look for one run**A look’s options: background colour, backdrop, lighting and the shape of the picture. They style this run only, and the sum for the batch is printed under the button. Local run, invented accounts; no image model was called.

Four views: a cost line, a cost confirmation, a spend warning and an hourly cap warning — **Cost in view before anything is spent**The sum under the button, a confirmation above 25 cents, a warning when a batch would pass the alert level, and another when it has more photos than the hourly cap allows. Local run, invented accounts; no image model was called.

The Cutout look selected, then the original photo beside the cut-out flowers on a checkerboard — **A real cutout**The Cutout look removes the background: the result is shown on a checkerboard and downloads as a transparent file. The pictures are the samples that ship with the app. Local run, invented accounts; no image model was called.

The Exact color option with its explanation open, and a custom colour picker — **The exact colour**Generated backgrounds drift a little from the colour asked for. Exact color rewrites the background to that colour after generation, and any colour can be picked. Local run, invented accounts; no image model was called.

A batch in progress with one finished result and two spinners, then a failed photo with a retry button — **A batch, one result at a time**Each photo gets its own result as it finishes, and a photo that fails can be retried alone. The photos here are drawings and the result tile is marked as a placeholder: no model was called. Local run, invented accounts; no image model was called.

A History page with a list of image results and one opened to show before, after and its settings — **Every result is kept**History keeps each result with who made it, the look and the settings used. The people are invented; the result is the sample that ships with the look. Local run, invented accounts; no image model was called.

Preset Creation

Looks the team builds and shares.

A look is built from a test photo and optional example images. The app reads the examples into settings, shows the sentence each setting adds to the instruction, generates three samples and saves the look for everyone, with its earlier versions kept.

Preset editor with a test photo, three sample outputs and background settings — **A team look in the editor**A team look reopened for editing, with its test photo, three sample results and every setting beside the sentence it adds to the prompt. An existing capture of the live product. Local run, invented accounts; the first picture is an existing capture of the live product.

Four steps of creating a look: inputs, a reading state, the editor and a retry for one failed sample — **From examples to a look**An example image and a test photo go in, the app reads them, and the editor opens with the settings it found. A failed sample can be retried alone, and a look cannot be saved until all three exist. The photos are drawings, the reading replays the project’s own test values, and the finished samples are tiles marked as placeholders. Local run, invented accounts; the first picture is an existing capture of the live product.

A look’s detail page beside the look library, and a version history with two rows — **Saved for the team, with its history**A team-made look’s page with its facts, test photo and results, and its version history after an edit and a revert. The result pictures are samples that ship with the app. Local run, invented accounts; the first picture is an existing capture of the live product.

For administrators

Spending is set by the team, and applies at once.

Administrators set a cap per person, a team budget, a warning level and the price booked per image. A saved change applies to the very next generation, and every change is logged.

An Image limits panel after a save, and the Generate area reflecting the new limits — **A limit changed, and felt at once**An administrator raises the cap and the warning level; straight afterwards the same batch in Image Studio no longer trips the cap warning. Local run; every figure and account is invented.

A Usage and Cost page with a total and two spend cards with line charts — **What image generation costs**The Usage page: total spend, image generation over time with its image count, beside the customer-service drafts. Every figure and account is invented. Local run; every figure and account is invented.

Measured, not assumed

The same flowers, on the colour that was asked for.

Two things have to hold for a result to be usable: the flowers stay the same, and the background is the colour that was asked for. The samples that ship with each look show the first; a script measures the second.

Before and after: red roses in a white vase on a warm backdrop, then on flat light gray — **A built-in look**A built-in look: the arrangement and vase stay, the background becomes a flat light gray. Bundled sample, not a new generation. Samples bundled with the app; nothing was generated for these pictures.

Two rows of four pictures: a single rose and its three results, a close-up of roses and its three results — **Two more looks**The Stem and Close-Up looks: each test photo beside the three samples that ship with the app. The close-up is recomposed into a rounded bunch, the limit the pilot found for tight close-ups. Samples bundled with the app; nothing was generated for these pictures.

Terminal output of a background drift report listing fifteen images with measured colours and verdicts — **Background colour, measured**The app’s own script over the fifteen bundled samples: the colour asked for, the colour measured, the drift, and the drift after the exact-colour correction. Three miss because their backgrounds are not flat. Real output, read from the bundled files. Samples bundled with the app; nothing was generated for these pictures.

Where it started · April 2026

A pilot that made the cost case first.

Pilot · April 2026

Before Image Studio there was a pilot: the same job done by calling image models directly, five models compared on one flower, and five catalog products restaged on five backgrounds. Its site had a working background studio, a research page and a playground. The pictures of the site are its own code running on my machine, and the result pictures are the pilot’s own saved outputs from April 2026.

Grid of catalog flower photos re-rendered on five different backgrounds — **Five products on five backgrounds**April 2026 pilot: catalog photos (top row) on five generated backgrounds. The leaf kept its shape every time; most tight flower close-ups were recomposed rather than kept, a limit the pilot recorded. Local run of the pilot site; the results are the pilot’s own saved April 2026 outputs.

A catalog photo of white anemones beside five generated versions labelled by model — **Five models, one flower**The catalog photo and what five image models returned for the same written description, from the pilot’s saved outputs. Local run of the pilot site; the results are the pilot’s own saved April 2026 outputs.

A dark page titled Product Background Studio with a drop area and style choices — **The pilot’s studio**Upload photos and choose a style: a plain white background, a company preset or a custom prompt. Local run of the pilot site; the results are the pilot’s own saved April 2026 outputs.

A catalog photo of white anemones beside a single anemone on a white background — **Original and result**The pilot studio’s result view: a catalog photo beside the pilot’s own saved result for it on a white background. Local run of the pilot site; the results are the pilot’s own saved April 2026 outputs.

A dark page titled AI Creative Studio with four tabs and a background swap panel — **A playground for four kinds of content**Background swap, photo creator, video ads and story videos, with an estimated cost for the batch. Local run of the pilot site; the results are the pilot’s own saved April 2026 outputs.

Engineering choices

The decisions behind the interface.

Call the models directly.

The pilot compared a credit-based tool with direct model calls on the same task: about 15 cents against an estimated 4 cents an image on the model tested. The released studio uses a newer, more expensive model, so the saving per image is smaller than the pilot’s estimate.