Mascot Cat on gpt-image: One Reference Keeps the Look, Pinterest Delivered 1.84 Thousand Impressions
A St. Petersburg finance company has spent two months drawing a ginger cat mascot with gpt-image, and it struggles with one thing: making the hundredth image show the same cat as the first. A text description produces a slightly different face every time, so the team attaches a single sample image (a reference) to every request. Over 30 days Pinterest delivered 1.84 thousand impressions, almost all of them on the entertaining pins rather than the financial ones. The takeaway: lock in a reference and check every image by eye.
AI-processed from Habr AI; edited by Hamidun News
A generative model can draw one nice picture. It gets harder when you need the same character on a hundred images for dozens of platforms: the hundredth must show the same cat as the first, and it must contain nothing that would make a moderator take the post down. A small finance company from St.
Petersburg has spent two months running a ginger fold-eared cat in a blazer this way. He has appeared on article covers, in Pinterest cards, on phone wallpapers and in videos. The gpt-image model draws him through the API, and no human illustrators are involved in the process.
Imagine asking a designer to "draw a ginger cat" a hundred times in a row without a photo of the hero: you would get a new face every time. So the team attaches a reference to every request, meaning the same sample image each time. Below is what has worked for them, what still breaks, and why Pinterest brought in only 1.
84 thousand impressions in 30 days.
How to keep the same cat
First finding: a text description of the character does not stabilize him. "A ginger fold-eared cat" comes out slightly different every time, with changes in face shape, shade and fur length. The only thing that helps is a reference image sent with every request.
The setup is as follows. There is one main reference that stays unchanged for months. The prompt contains a phrase like "the same ginger Scottish Fold cat from the reference", and the cat's appearance is not described again. Clothing is specified separately and changes with the scene: blazer, sweater, hard hat, overalls. The face and coat color come from the reference.
The prompt is assembled from the same blocks in the same order: style, character, action, location, prohibitions. It is better to choose an action with an object in the cat's paws: scenes where the cat holds something or struggles with something come out livelier than portraits. A recognizable city detail appears in the window, for example the dome of St. Isaac's Cathedral, which the model renders reliably.
Tasks are stored in JSON, and a script runs them in parallel, two variants per task. Before launching, the script compares the model's per-image price against a threshold and refuses to run if the price is higher. This is insurance in case the aggregator suddenly changes its tariff while forty images are in the queue.
What breaks
Text. The phrase "no text" works poorly. On papers, laptop screens, calendars and signs, the model still draws letters, usually gibberish, and it immediately gives away that the image is generated. Describing the text as unreadable works better: "text unreadable, blurred", "abstract colorful code lines, unreadable", "chalkboard with drawings only". The model draws a blurred patch more willingly than an empty space.
If text is needed, for example in infographics for Pinterest, the team used to build cards in HTML, and they looked like a template from a website builder. Now the model lays out the whole card and writes the Russian text itself. The prompt demands that it reproduce the text strictly verbatim, with no extra words and no Latin letters, each item in quotation marks. The results are better, but typos still happen: "подтинуть" instead of "подтянуть", "раз месяц" without the preposition. So every variant is checked by eye, and if both variants have errors, the task is regenerated. The text is kept short, 2–5 words per item.
Paws. The more complex the action, the higher the chance of a third paw. For example, when the cat holds a phone and signs a document at the same time. The fix is to simplify the scene.
Logos and other people's characters. If a scene includes "Gosuslugi" or a "banking app", the model may draw something resembling a real logo, and the image starts to look like an ad or a forgery of someone else's brand. The team writes "abstract banking app screen with a big green checkmark". For a scene based on "The Night Watch", the cat in a feathered hat and boots came out too similar to a famous cartoon Puss in Boots, and the image was thrown out. Now, for scenes based on paintings, the team uses only public-domain works and does not give the character attributes of recognizable heroes. A stray mark in a corner is simply cropped out.
Platform rules and Pinterest numbers
One image does not fit everywhere. The team has built up a list. On "T-Zh", the first image becomes the cover, and it must contain no text or logos.
On "Klerk", the cover also has no logos or text, and its size is 2048×1152. Drive2 users like realism, so the team posts photorealistic scenes there with the caption "illustration generated". On Pinterest, generated pins are labeled as AI-modified.
On Pikabu and in bank forums, a mascot cat looks like advertising, and simple two-panel sketch memes without the character work there. A cover is never a table or an infographic: a table on the first screen looks like a report, and infographics live only inside the text. The team does not strip provenance metadata (C2PA) from files and does not pass the images off as photos.
There are three sizes: 1536×1024 for covers (then cropped to the required ratio), 1024×1536 for pins, and 1088×1920 for 9:16 wallpapers, where the team asks the model to keep the top fifth of the frame calm for the clock.
Pinterest gave the most unexpected result. In the first 30 days the account got 1.84 thousand impressions, and almost all of them went to three entertainment pins: a paper-doll cat (697 impressions), clothes for it (219) and a coloring page (195). The financial infographics, the reason the whole project was started, got 4–10 impressions each. The character attracts attention where he is interesting in his own right, while financial advice with a cat is still just financial advice. There are now more entertaining images, and financial topics remain in the minority.
What to take away
Before publishing, the team checks six things: whether the cat looks like the reference (face, coat color, ears), whether there are extra paws or fingers, whether there is readable gibberish on small objects, whether there are other people's logos or resemblance to well-known characters, whether the text matches the prompt verbatim, and whether the image fits the platform's rules. The Pinterest numbers come from one account over one month, so it is too early to generalize from them. The practical takeaway does not suffer from this: if you run a character or brand on generated images, keep one unchanged reference, describe text in images as blurred, and leave a human for the final check.
Generation removes the illustrator, but not the editor: something that cannot be published turns up regularly.
- Apr 22, 2026Prompt Decomposition in Gemini and Kling: How to Recreate a Pinterest Look in Your Brand Style
- Nov 18, 2026KandiShop raises AI face similarity to the original from 52% to 75% without fine-tuning
- Oct 29, 2026Black Forest Labs releases FLUX.3 Klein: 1024×1024 in 2 seconds on an H100
- Nov 8, 2026OpenAI Releases GPT-Image 2.6: Two Models, GIF and SVG Export via API
Key Questions Answered
How to keep the same cat?
First finding: a text description of the character does not stabilize him. "A ginger fold-eared cat" comes out slightly different every time, with changes in face shape, shade and fur length. The only thing that helps is a reference image sent with every request. The setup is as follows. There is one main reference that stays unchanged for months. The prompt contains a phrase like "the same ginger Scottish Fold cat from the reference", and the cat's appearance is not described again.
What breaks?
Text. The phrase "no text" works poorly. On papers, laptop screens, calendars and signs, the model still draws letters, usually gibberish, and it immediately gives away that the image is generated. Describing the text as unreadable works better: "text unreadable, blurred", "abstract colorful code lines, unreadable", "chalkboard with drawings only". The model draws a blurred patch more willingly than an empty space.
What are the key details on Platform rules and Pinterest numbers?
One image does not fit everywhere. The team has built up a list. On "T-Zh", the first image becomes the cover, and it must contain no text or logos. On "Klerk", the cover also has no logos or text, and its size is 2048×1152. Drive2 users like realism, so the team posts photorealistic scenes there with the caption "illustration generated". On Pinterest, generated pins are labeled as AI-modified.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.