AI image + video workspace · 2026Creative direction should be visible before generation.
Mize turns loose intent, references, and visual judgment into a direction people can see, compare, edit, and confirm before a model executes it.
- Project
- Independent concept
- My role
- Product definition, UX/UI, prototype
- Scope
- Home, Workspace, Library, Assets
The complete product
From the first idea to the work you keep.
Seven connected pages support the whole creative journey. Choose a page below, then explore the working prototype inside it.
A starting point for an unfinished idea.
Start with a creative goal, explore what the product can do, or go straight into the workspace.
The premise
The prompt was carrying too much.
Most generative tools compress creative direction into one text field. Style, lighting, material, composition, and camera intent disappear into prose, so people repeatedly rewrite prompts while guessing what the model understood.
I reframed the prompt as the final output of a decision-making process—not the place where every decision must begin.
Product thesis“Prompt is the final output of a decision-making process, not the entry point.”
Interaction principles
A system for visible judgment.
Keep language flexible
The brief carries purpose, feeling, and constraints without becoming a rigid prompt formula.
Make decisions visible
High-confidence visual choices become inspectable chips with evidence, category, and provenance.
Let AI propose
AI can create meaningfully different routes, but it cannot silently change the current direction.
Protect confirmation
Every accepted direction becomes a recoverable version whose confirmed choices survive reruns.
Direction + prompt
The direction sits above the prompt for a reason.
Current direction gives visual decisions a persistent, editable place. The prompt underneath carries the subject, purpose, feeling, and constraints that still need language.
01Current direction
02Prompt · natural language
Create a clear-glass botanical fragrance campaign for a spring launch. Keep it fresh, editorial and believable, with enough negative space for a logo.
Decisions you can inspect and edit
Each chip names a visual choice. Hover reveals an example, category, explanation, and Library source; remove a chip or add a Library decision without rewriting the brief.
Language with room for intent
The prompt explains what to make and why. In the prototype, entering an exact Library term turns it into an editable chip; open-ended or ambiguous language remains in the text field.
One baseline across both modes
Normal Mode uses Current direction for direct generation. AI Mode labels the same choices Confirmed decisions: suggested routes stay separate until the user confirms a version.
Core workflow
AI suggests a change. The user confirms a version.
AI Mode lives inside the same workspace as direct generation. It behaves as a direction assistant, not an always-on chat or an autonomous art director.
- 01 · Baseline
- 02 · Compare
- 03 · Review
- 04 · Confirm
Set the protected baseline
References, a natural-language brief, and confirmed decisions establish the state AI must respect before it suggests anything.
Directions are diffs, not snapshots.
The labels under a direction card describe what would change relative to the current baseline. A selection is only a proposal under review; Current direction stays untouched until confirmation.
Protected choices
Sunlit Botanical
- Keep
- Photographic Editorial · Clear Glass
- Add
- + Botanical Setting
- Replace
- Natural Daylight → Hard Sun
Centered Product → Overhead Still Life
New current direction
Preview versus execution
A route preview is not a finished result.
The preview rehearses the proposed visual logic. Only after confirmation does Mize execute a new result from the updated direction, so the user can distinguish exploration from generation.
The knowledge layer
Library is evidence, not inspiration wallpaper.
Each item names one reusable creative decision and proves it with three or four examples. The subject may change; the decision must remain legible across the evidence set.
Browse a decision, expand its proof, then bring it into the active workspace.
The interaction preserves the distinction between a reusable knowledge item and one concrete example, so users can reference either the rule or the execution.
Image Library
One decision. Several ways to see it.
The cursor opens each image stack into an evidence set. Compare subjects, lighting, and composition to understand which visual choice stays consistent.
Video Library
A vocabulary you can watch in motion.
The same hover unfolds moving examples side by side. Camera movement, a focused reveal, and process rhythm make each motion decision visible across different scenes.
One shell, two modalities
One workspace for still images and motion.
Image and video use one workspace structure, with separate model lists and result sets. The Video Library organizes camera motion, subject motion, pacing, and shot structure as a vocabulary for directing movement.
- Image
- Lighting · Material · Composition · Color
- Video
- Camera motion · Subject motion · Pacing · Shot structure
Information architecture
Knowledge, ownership, and active context stay separate.
System knowledge
Reusable visual decisions supported by curated evidence and shared vocabulary.
Save or add to WorkspaceUser-owned material
Uploads, generated series, style references, characters, products, and visual systems.
Manage, delete, or reuseKnowledge bookmarks
A lightweight way to return to Library items without copying them into owned Assets.
Always synced to LibraryActive creation context
The brief, references, confirmed direction, versions, model settings, and results for one task.
Protected across rerunsWhat the iteration clarified
Control needed to be a state model, not a slogan.
- 01
Separated Normal Mode’s direct Generate path from optional AI exploration.
- 02
Changed AI direction cards from complete snapshots into explicit proposed diffs.
- 03
Made selection visibly different from confirmation and preserved the previous version.
- 04
Separated Library knowledge, Saved bookmarks, and user-owned Assets.
- 05
Rebuilt the representative task so references, decisions, previews, and results tell one coherent story.
Current outcome
A working front-end prototype for testing the product logic.
The prototype covers Home, Image and Video Workspace, AI direction exploration, direct generation, Library evidence, Assets, Saved, recent workspaces, recovery, and model selection. Model calls and persistence remain simulated so the interaction model can be evaluated before backend investment.
Back to selected work