LUCID
PAST
Creating Virtual Dreams from Historical Memory
The Problem Space
Archives as Unexplored Territory
The Library of Congress holds over 16 million images in its Prints and Photographs Division, with around 1.6 million digitized. Museums worldwide have similar collections. Yet most remain locked behind keyword search and chronological browsing - interactions designed for librarians, not storytellers.
XR enables natural, curiosity-driven exploration. But existing museum VR experiences are linear guided tours - just digital versions of audio guides.
Three core issues I identified
Discovery is broken.
Context is rigid.
Engagement is shallow.
The Insight
The brain holds personal memory.
Archives hold collective memory.
Dreams are how we explore one.
LucidPast is how we explore the other.
When you dream, your brain does not present memories in chronological order. It follows associations - a face leads to a place leads to a feeling. The result is not chaos. It is meaning, arrived at sideways.
Archives work the same way. The Library of Congress does not know which photograph will matter to you, or why. Keyword search forces you to already know the answer before you begin. LucidPast replaces the search bar with the dream - letting association, attention, and curiosity do what they have always done best.
Core Innovation
How Dreams Actually Work
I asked: Can we recreate this algorithmically?
The system I designed:
- →
Gaze reveals intent
- →
Intent triggers transitions
- →
Transitions build themes
- →
Themes create meaning
Design System Architecture
I broke the experience into three interconnected subsystems. Each solves a specific design challenge.
Persistent Object Transition
Environmental changes in VR cause disorientation. How do you morph between completely different scenes without losing the user?
The gazed object becomes a spatial anchor - context morphs around it, eliminating disorientation.
Three-Act Progressive Narrowing
Pure random exploration becomes meaningless. Pure curation removes agency. How do you balance both?
Three-act structure applied algorithmically - 100% open for 7 minutes, narrowing to 30% of options by minute 20.
Pathway-Dependent Interpretation
The same photograph can reveal labor history, tech innovation, or cultural change. Institutions usually pick one.
Pathway priming activates different semantic networks - the same photograph means different things depending on how you arrived at it.
Interaction storyboard - 36 panels mapping the full experience flow before implementation.
| Phase | Duration | System Behavior | User Experience |
|---|---|---|---|
| Act 1 | 0-7 min | Observes gaze patterns, 100% open | "I can go anywhere" |
| Act 2 | 7-13 min | Narrows to 40% of options matching detected themes | "This is getting interesting" |
| Act 3 | 13-20 min | Focuses to 30% of options, clear thematic thread | "Oh, this is about..." |
Interaction Design
Gaze as Primary Input
Why gaze? I evaluated five alternatives:
Dwell time calibration validation: Research shows 300-400ms causes accidental activation ("Midas Touch problem"). Apple Vision Pro's Dwell Control (accessibility mode) defaults to 1000ms - I validated this as the right threshold to avoid accidental activation.
Two Interaction Modes
Early concepting revealed a problem: "What if I'm looking at a rendering artifact or examining something out of confusion, not interest?" Solution: Users choose at entry, switch anytime.
Flow Mode
- For maximum immersion
- Pure 1-second gaze → automatic transition
- Minimizes cognitive load
Intentional Mode
- 1-second gaze + pinch gesture confirmation
- Explicit gestures don't break flow state
Technical Implementation
Volumetric Reconstruction: Two Approaches Validated
Challenge: Creating explorable 3D environments from single flat photographs without fabricating content beyond what's visible. I validated two fundamentally different pipelines on the same test dataset of 10 diverse archival photographs:
Model Selection - Iterative Validation
Pipeline A — Depth Pro + Mesh
Depth Pro depth estimation
~8s/image
Point cloud generation
~2 min/image
Poisson surface reconstruction
~15 min/image
Texture mapping
~3 min/image
Total:~20 minutes per image
Output:textured polygon mesh (.glb)
Pipeline B — Apple SHARP Splatting
SHARP inference
<1s/image
Total:under 1 second per image
Output:Gaussian Splat (.ply)
Why SHARP produces better results for LucidPast:
Test dataset validated across extreme conditions:
Both pipelines generated convincing 3D across all conditions. SHARP showed measurably superior texture preservation in portrait-heavy content - the dominant image type in institutional archives.
WorldLabs (marble.worldlabs.ai) - Tested
Single-photo to 360-environment conversion. Tested with Migrant Mother - the consistent benchmark image across all pipeline validation. WorldLabs uses AI image generation to approximate and fill invisible areas beyond the camera frame.
Result: Migrant Mother's face was completely altered and unrecognisable. The child was removed entirely. The background was fabricated as open ground with multiple trees - approximated from the partial background visible in the original.
The confirmed constraint
The moment a system generates content beyond the photograph, it loses the historical subject. This is not a technical failure - it is a category error. You are no longer showing the archive. You are showing an AI's interpretation of the archive.
LucidPast limits parallax to visible content only. What the photographer did not capture, the system does not invent.
WorldLabs output - Migrant Mother's face altered beyond recognition, child removed, background fabricated from partial scene data.
The Narrative Sequence
2D Slideshow Validation
I built a functional slideshow with simulated gaze tracking to validate narrative mechanics before investing in XR implementation. Internal testing showed theme emerged clearly by image 7-8 despite no explicit narration.
Migrant Mother (1936)
Textile mill child laborer
Ford assembly line workers
Newsboys selling papers
Depression breadline
Store window display
Mannequins with artificial happiness
Beauty cream advertisement
Makeup application behind-the-scenes
Commodification of authenticity—society manufacturing happiness during Depression while real suffering exists steps away.
Emergent theme confirmed by image 7-8 without explicit narration. Pivot point felt surprising yet inevitable - the "dream logic" quality I was targeting.
Key Design Decisions
User Validation & Testing
This section will be completed after conducting user testing with 3-5 participants following Nielsen Norman Group's qualitative testing guidelines. Content will include testing methodology, key findings on thematic emergence rates, narrative coherence ratings, and representative participant quotes.
Outcomes & Reflection
What This Project Validated
Design Hypothesis
Attention-driven sequencing generates emergent themes. Progressive narrowing balances spontaneity with coherence.
Technical Feasibility
Two volumetric pipelines validated. SHARP produces photorealistic splats in <1s, enforcing ethical constraints architecturally.
Interaction Pattern
Persistent object transitions reduce disorientation. Two-mode gaze+gesture system addresses false-positives without breaking flow.
Skills Demonstrated
Interaction Design
Systems Thinking
Technical & Research
What I Learned
Making history feel like memory
Image Sources
- Archival photographs: Library of Congress, Prints and Photographs Division
- Neil Armstrong portrait: NASA (public domain)
- Black Mirror - Eulogy (S07E05): Netflix / House of Tomorrow - used for critical commentary
- Depth map outputs and 3D reconstructions: author's own work