Atlabs AI · Case study 01

Create Image

Designing one calm surface for 24 fast-moving AI models. As of October 2026, ~6,800 creators have used it to make ~55,000 images.

Headlines, pictures and numbers only

RoleFounding Designer, sole designer. End-to-end: research, flows, UI, design system, copy.

ScopeCreate Image + the shared "mini app" shell later reused across Edit, Inpaint, Reframe, Upscale, Image to Video, Lip Sync and Motion Control.

The image detail view: an illustrated hiker on a forest trail, and beside it the prompt, model, dimension, format, resolution and credits, with Reuse Inputs and six next-step actions below — One image, its full recipe, and six ways to keep going.

As of 24 Oct 2026 · external users only

~55kimages requested

~6.8kcreators have used it

70%of requests use a reference image

2.6×more monthly active users, Oct 2025 → Oct 2026

TL;DR

Atlabs needed a direct way for creators to use the latest image models without going through a full video workflow. I designed Create Image from scratch, and with it a left-input / right-output shell that became the template for every Atlabs mini app. It grew into a core tool for a committed group of creators, who mostly use it to refine reference images into high-resolution, video-ready frames; the data also showed me where it still loses people, which sets the next round of work.

Context

Atlabs turns ideas into animated videos. Inside a story, images are generated for you. But creators kept asking for direct control: make a character sheet, fix one frame, test the model everyone was talking about that week.

There was nothing to iterate on. No previous version, no pattern to extend. I designed Create Image from zero, and treated it as the first instance of a system rather than a one-off page, because more tools were clearly coming.

The problem

Three tensions shaped every decision

The model list never stops changing.

The picker now holds 24 models from 9 families: Nano Banana, Seedream, GPT Image, Flux, Grok Imagine, Qwen, Recraft, Runway, Ideogram. Model names say nothing useful to most creators.

Answered by

Generation is iterative, but most tools treat each output as final.

The useful move is changing one variable and running again, not starting over.

Answered by

Credits make every click feel expensive.

Generations take a median of 32 seconds and, as of October 2026, cost from 0.6 to 12 credits depending on model, resolution and quality. Uncertainty about either makes people hesitate or feel tricked.

Answered by

So I set four rules

Same shell, different job. Learn one mini app and you know them all.
Decide by trade-off, not by name. Show what a model is good for, not what it's called.
No surprises. Cost and wait time are visible before you commit.
Every output is an input. A result should be one click away from its next step.

Key decisions

Seven decisions, each with the evidence we saw afterwards. The data shows behaviour, not why, so it's described as consistent with each decision, not caused by it.

Decision 01 Same shell, different job

Tools and models as two doors into the same room.

Task-first and model-first people land in the same tool, so new models never change the navigation.

The Image and Video menus split into Tools (what you want to do) and Models (what you want to use). Most people arrive task-first ("I need to upscale this"). Power users arrive model-first ("I want to try Seedream 5 Pro"). Both doors lead to the same mini app, so the model list can grow every week without the navigation changing shape.

What we saw

model families in active use, and the long tail (Flux, Grok Imagine and others) still gets real traffic.

The Image and Video menus side by side. Each splits into a Tools column (Create Image, Edit, Inpaint, Reframe, Upscale, Lip Sync) and a Models column (Nano Banana, GPT Image, Seedream and more) — Two doors into the same tool: what you want to do, or what you want to use.

The Image menu: a Tools column and a Models column — Image menu.

The Video menu: a Tools column and a Models column — Video menu. Two doors into the same tool: what you want to do, or what you want to use.

Decision 02 Same shell, different job

One shell for every mini app.

One layout, inputs left and results right, now carries every mini app.

Inputs on the left in the order people think: model, prompt, references, then output settings. Results on the right. I designed this as a system from the start, and it now carries every mini app on Atlabs. New tools ship faster, and users never relearn the layout.

What we saw

tools now run on the same shell, from Create Image to Upscale, Reframe, Lip Sync and video.

Create Image: model, prompt and reference inputs on the left, a three-step How it works guide on the right — The shell every mini app now shares. Switch tools: inputs stay left, in the order people think, and results stay right.

Generate Video: model, frames and prompt inputs on the left, a How it works guide on the right — The shell every mini app now shares. Switch tools: inputs stay left, in the order people think, and results stay right.

Decision 03 Decide by trade-off, not by name

Choose a model by speed and resolution.

Every model shows wait time and max resolution: the two things people ask before generating.

The model picker leads with a curated Featured list and has search. Each row shows two facts: estimated time (~9s to ~45s) and maximum resolution (up to 1K / 2K / 4K). These are the two questions creators actually ask before generating: "how long will I wait?" and "is it sharp enough for my video?". NEW and PREMIUM badges flag what just landed and what costs more.

What we saw

54%

of requests are at 2K or higher, and the most-used model family is a premium one, not the cheapest. Creators pay for quality when they can see what they're getting.

The model picker: a search field, then a Featured models list where each row shows an estimated time, a maximum resolution, and NEW or PREMIUM badges — Each model shows the two things people ask before generating: how long, and how sharp.

Decision 04 Consistency over novelty

References are first-class, not an afterthought.

References sit right under the prompt, because video needs the same character frame after frame.

Reference images sit directly under the prompt, with room for up to 16. In an AI video product, consistency matters more than novelty: the same character, the same style, frame after frame.

What we saw

70%

of requests include at least one reference image. In practice Create Image is mostly an image-to-image and editing tool, not a blank-canvas one.

The model field, the prompt box and, directly under it, the Reference Images area with room for 16 images — References sit right under the prompt, because consistency matters more than novelty in video.

How people actually use it % of external requests, as of 24 Oct 2026

Use a reference image70.6%

2K resolution or higher53.7%

16:944.4%

9:16 vertical30.6%

4K11.0%

More than one output1.4%

Decision 05 No surprises

Cost lives inside the button.

The price is inside the button and updates live, so nothing surprises you.

The credit cost is part of the Generate button itself and updates as model, resolution, quality and outputs change. There is no separate pricing step and no surprise afterwards. Try the panel: these are the real prices.

What we saw

1.4%

of requests ask for more than one output. People generate deliberately, one image at a time, which is the behaviour visible cost encourages.

Try itReal prices from the app, as of 24 Oct 2026

Model

Resolution

Quality

Outputs

~19s estimated

A working copy of the settings panel. Change anything and the cost inside the button updates, as it does in the app.

Decision 06 Every output is an input

History built for iteration: Reuse Inputs.

Reuse Inputs reloads a whole recipe in one click, so you change one thing and run again.

History has three views for three jobs: grid to scan, filmstrip to compare variations large, list to read prompts next to results. Hovering any image shows its full recipe: prompt, model, ratio, resolution, reference count, credits, time. Reuse Inputs reloads the whole recipe into the panel in one click, so the next run changes one variable instead of rebuilding everything by copy-paste.

What we saw

332

characters is the median prompt; the average is ~110 words and the longest passes 34,000 characters. Prompts are too long and too carefully built to retype.

History in grid view: generated images in a three-column grid — Three views for three jobs: scan, compare, read.

History in filmstrip view: one large image with a strip of variations beside it — Three views for three jobs: scan, compare, read.

Hover card on a past image showing its prompt, chips for model, ratio, resolution, outputs and references, its credits and time, and an orange Reuse Inputs button — Every image keeps its recipe. Reuse Inputs reloads it in one click.

Try itThe iteration loop in three steps

A past generation: an illustrated hiker on a forest trail above a river valley

animated scene of a beautiful hike

Nano Banana Pro16:94K1 output

⛁ 5 · 37s

Reuse inputs
Change one thing
Generate

Model—

Prompt—

Dimension—

Resolution—

Generate⛁ —

Decision 07 Every output is an input

Every image leads somewhere.

Every result offers its next step: animate, edit, upscale, reframe, lip sync.

The detail view turns each output into a starting point: Image to Video, Edit, Upscale, Reframe, Add Lip Sync, Motion Control. The mini apps stop being separate utilities and become one pipeline: image → video.

What we saw

31%

of requests are vertical 9:16, a format that only makes sense for short-form video. People are making frames, not pictures.

Six action buttons under an image: Image to Video, Edit, Upscale, Reframe, Add Lip Sync and Motion Control — Every output is an input to the next tool.

A small detail

The empty state teaches the three steps with one cupcake that gets more refined at each step: write → guide with a reference → fine-tune.

How it works: three cards showing one cupcake image getting more refined at each step: write a prompt, guide with a reference, fine-tune — One example, three states, no wall of text.

Impact

Since launch in September 2025, to 24 Oct 2026. External users only, rounded.

~55,000 images requested by ~6,800 creators.
Monthly active users grew ~2.6× from October 2025 to October 2026, and October 2026 was already the busiest month for requests with a week still to go.
A real core of committed users: a small group of paying creators makes most of the requests. For them, Create Image is a daily production tool.
The model system scales: creators actively use 9 model families, and new ones get picked up within months of landing.
Agents use it too: the same models are available through Atlabs MCP, and AI agents now send requests as well as people.

What the data told me that I didn't want to hear

Most people who tried Create Image made one image and never came back.

The tool works well for people who already know what to type. It does far less for someone facing an empty box that says "Enter here". Power users bring long prompts and reference images; newcomers bring a vague idea. The interface treats them the same.

I'd rather show this than hide it. It is the clearest signal for what to design next.

What's next

Three bets, all aimed at the first five minutes:

Better prompt writing. An "enhance" action that turns a rough idea into a detailed prompt, tuned to the selected model's strengths.
Suggested prompts. Starters based on common Atlabs use cases (character sheets, product shots, scene backgrounds), so nobody starts from a blank box.
An inspiration gallery. Example images with their prompts attached. Clicking one runs the same Reuse Inputs pattern, so discovery and iteration become one mechanic.

How I'll know it worked: a higher share of new users making a second image, and more users returning on a second day.

Reflection

Designing the shell first paid off. Every new tool since has been cheaper to design and build because the hard decisions were already made once.
Show the decision inputs, not the catalogue. Speed, resolution and cost turned out to matter more than model descriptions.
Data is a design input, not a trophy. The one-and-done finding is the most useful line in this case study.

Footnotes

Source: Atlabs production data, pulled 24 Oct 2026. Figures are rounded and exclude internal staff and company accounts unless stated.
Monthly active users include a small number of internal accounts, tiny next to the monthly total. Request totals are not shown by month because internal usage inflates them.
"Requests" counts generation attempts in the standalone Create Image tool, including edits that re-enter it via Edit Image. Image panels inside the story editor are not included.
Behavioural patterns are described as consistent with design decisions, not caused by them. Prices in the interactive panel are the app's prices on 24 Oct 2026.

Get to know me.

The story behind the work, or my résumé on one page.

my story

How an engineer from IIT Roorkee ended up designing AI video tools.

Read it →

My story · 2 min read

Rupam

Founding Designer · Atlabs AI