case study · Multimodal AI · Gemini Hackathon

ImageFuse

Merge elements from one image into another with a sentence.

Role
Solo: design and engineering
Timeline
2026
Status
Paused· Gemini credits exhausted
Links
live ↗code ↗demo ↗
Stack
Next.jsTypeScriptGeminiCanvasTailwind

Upload two images, describe which elements to combine, and Gemini plans the edit while an image model renders the fused result.

01

The problem

Two fashion designs needed to become one visual concept. Doing it properly meant Photoshop skills and hours of masking; doing it fast meant a bad collage.

Who it’s for

Designers, small fashion brands and anyone who thinks in “this part of that, on this”.

02

Goals & constraints

Goals

  • Describe the merge in words, not masks.
  • Keep the result realistic.
  • Simple enough to use in one sitting.

Constraints

  • Reasoning models return text, not images
  • Response time has to stay reasonable

03

Approach

  1. 1.

    Two-model pipeline

    A Gemini reasoning model reads both images and the instruction and writes a precise edit plan. A separate image model executes it. Each step has its own configurable fallbacks.

  2. 2.

    Conversation over controls

    Two uploads and a sentence. The complexity lives in the plan, not the UI.

04

Key decisions

Plan, then render

Splitting reasoning from generation made results far more faithful to the instruction than one model doing both.

!

Challenges

Prompt clarity for visual reasoning

Vague instructions produced vague edits; the plan step rewrites them into specific, region-level instructions.

Different image structures

Inputs vary wildly in framing and lighting, which the plan has to account for.

✓

Outcome & lessons

  • Submitted to the Gemini hackathon; live instance paused while API credits are out.

What I learned

  • Multimodal workflows
  • Designing clean UX around unpredictable AI

What’s next

  • Region selection on canvas as an optional precision mode
next projectExamDrill→