Internal creative production tool 2026

BLINKCHARACTERSTUDIO

How do you keep an AI-generated character recognisable from one shot to the next?

I designed a locally runnable character-reference tool that turns a loosely defined production problem into a structured, reviewable workflow. A user defines or uploads a character, approves a visual reference, and builds a reusable package for subsequent CGI and AI-assisted shots.

See the application demo
01 /

See the tool in action

Actual application recording
32-second walkthrough

A short walkthrough of the working interface, from character setup to reviewing a completed reference board.

Actual application screen recording32 seconds · setup and reference-board review
02 /

The problem behind the tool

The production problem

A character may look convincing in one image and drift in the next. Face, clothing, shoes and small accessories are especially vulnerable when references are recreated from memory or prompts.

The design challenge

How can a creative team create, inspect and preserve one agreed character identity without depending on a fragile sequence of individual prompts?

The response

A guided character specification, visible checkpoints, approved image references and a consistent export format. The human decides what is acceptable; the system manages the stages.

03 /

The interaction is the safeguard

Two entry points · one controlled production path
A / DEFINE

Start from a specification

Set the character's appearance, wardrobe, equipment and project context. Generate a preview only when the user chooses to proceed.

OR
B / REFERENCE

Start from an existing photo

Upload and crop a real reference. A full-body image can serve as the preview directly; a partial image may first need a generated full-body preview.

01Define / upload
02Preview
03Human approval
04Face & detail references
05Turnaround
064K board & export
A deliberate human checkpoint

The application does not pretend to verify identity.

A person reviews the preview, checks face and clothing details, and explicitly approves or rejects it. The approved asset and specification are preserved for subsequent steps. The completed board is reviewed again before acceptance.

04 /

The characters behind the workflow

Original character sheets
Selected production examples

11 original character sheets, presented as a continuous sequence. Each reference stays on screen for two seconds; pause or use the arrows to inspect individual boards.

1 / 112 seconds per sheet · click to enlarge
Portable output

Visual references + structured character record

The application is designed to export an assembled character board and a structured character package, keeping references and metadata available for downstream production.

05 /

What I designed and built

01 / Character specification

Editable character attributes, project metadata, garments and equipment, with an alternate photo-upload and crop route.

02 / Interaction & approval logic

Preview review, explicit approval or rejection, revision handling and a second acceptance checkpoint for the completed board.

03 / Multi-stage AI orchestration

Separate preview, face/reference and body-turnaround stages. Saved outputs and approval state determine what can happen next.

04 / Production-aware safeguards

Paid-call confirmations, per-character attempt limits, cancellation, saved-stage resume and no automatic chargeable retry after an uncertain failure.

05 / Reference-board assembly

Fixed 4K composition, adjustable panel crops, programme-set labels and colours sampled from source images, without an extra generation step.

06 / Character library & exports

Local records, named revisions, source files, model metadata and portable exports designed to support continuity between production shots.

06 /

Behind the interaction

Technical choices serve the production task

I approached the tool as a production system: minimise repeated decisions, make expensive actions intentional, and keep usable outputs even when a generation stage fails.

ReactNode.jsSQLiteSharpOpenAI / Gemini / Krea routesAI-assisted development
PROVIDER OPTIONS

Three adapters are implemented and mock-tested. A character stays linked to its chosen route once production begins.

LOCAL-FIRST

Character metadata is stored in SQLite; images are kept as files, not dependent on a third-party image library.

SPEND CONTROLS

Reservations and confirmations reduce accidental duplicate generation. They do not guarantee the provider's final invoice.

Prototype status and what has actually been verified

The supplied Windows v1.5.1 package reports 53 automated tests passed, zero failures, covering workflows, persistence, exports, safety checks and mock provider transport. The application builds locally and includes a launcher and test fixtures, which are deliberately excluded from this portfolio gallery.

Not yet verified in this build: live paid image generation, production image quality or identity consistency across outputs, browser interaction QA, and live company Google sign-in / Drive integration. The supplied screen recording and original character sheets show the interface and reference outputs. They do not establish a company-wide deployment or independently verified generation performance.

07 /

Why this project matters

What began as a recurring visual-production frustration became a tool with a user journey, approval gates, technical boundaries and a tangible output. For me, creative technology means using visual judgement and prototyping to make a difficult task easier for someone else to perform.