TunnelVision

Origins / Attribution

TunnelVision originated with filmmaker Terran Boylan, who developed the concept and manual filmmaking technique through work for his own YouTube channel and video content. The technique was later adapted for the Digging in the Dirt project, where we used the playful name TunnelTV. The research documented here extends Terran's original TunnelVision work into an agentic filmmaking system.

Original TunnelVision concept and manual technique — Terran Boylan
Agentic Director/Cinematographer architecture and Camotion research — Dan DeGeest

Current research baseline

Current product: TunnelVision is an AI filmmaking system for directing continuous journeys through imagined worlds. The working loop is SPECIFY → DIRECT → CONSTRUCT → SHOOT. CREATE JOURNEY is the Director action. Agency is DIRECTED or AGENT. A Project camera grammar (POV, FOLLOW, LEAD, or MOUNTED) applies to the whole journey. Nano Banana 2 at 2K is the current strong canonical configuration. Cinematographer writes a Segment Motion Plan from the actual pair; segment guidance is injected before the grammar-specific locomotion baseline. Adaptive Camotion conditions A′/B′ as pace × depth × destination protection × VP protection. Greenhouse and Gravity City movies show the pipeline working on spatially plausible and near-unshootable journeys. Visual checkpoint: Product Slice 8. Camera grammar: Product Slice 13.

Ghost Library / Seedance 2.5. Conservative directed A→B video baseline: Camotion 01.5, depth-banded outgoing exposure, strength 0.08. Current Camotion operator: 01.8 route-preserved exposure. 01.9–01.12 are completed still-only evidence and are not promoted. 01.12 found no useful pristine baked-exposure operating window (classification C + D) and closes the current exposure-operator tuning branch. Integration Test 01 — The Wardrobe Loop is the first unattended end-to-end movie experiment. A later A→B Seedance 2×2 (seed 70) supported retaining Camotion plus Cinematographer locomotion prompting as complementary Phase 1 pieces: Camotion is motion-state conditioning, not image enhancement. A three-pair scene-aware strength experiment preferred bounded per-canonical selection from {0.02, 0.04, 0.08} over fixed 0.08. Camotion Phase 1 is frozen. It is not Camotion 01.13. A later E→A shootability experiment showed the Cinematographer can reject an unshootable direct route, propose a spatial intermediate camera position, inspect the actual generated set, and stop before video if the geometry is not shootable. TunnelVision Research Phase 1 is complete. Product development has begun. Product Slice 1 introduced the fixture-driven Plan | Shoot filmmaking workspace. Product Slice 2 records Plan as conversation → storyboard; Shoot remains the Destinations / Journey production workspace. See Product Slice 2. Product Slice 3 adds a thin autonomous Director over that Plan shell. First live Director observation: Product Slice 3. After that, the filmmaker can replace A, plan from an arbitrary story, and construct destination B from A through image-conditioned edit. First construction observation: Director-derived destination construction. Sequential Construct from the preceding actual destination is now product behavior. Product Slice 4 records the current Plan | Shoot UI with Project-panel PLAN, sequential destination construction, BLOCK, and SHOOT. Visual checkpoint: Product Slice 4. Product Slice 5 adds the Project panel Video model and records the first mid-tier comparison takes. Visual checkpoint: Product Slice 5. Product Slice 6 records destination look-ahead: Construct injects the following beat's visual description as far-field only. Visual checkpoint: Product Slice 6. Product Slice 7 records an earlier Plan | Shoot UI checkpoint on a live ten-destination train journey: media facts on the selected still, header Debug and Directed/Autonomous, MOTION Plan / FOOTAGE Generate, and stored Camotion overlay on destination preview. Visual checkpoint: Product Slice 7. Product Slice 8 is the current product overview: CREATE JOURNEY, DIRECTED / AGENT, SHOOT / RESHOOT, adaptive Camotion, and assembled greenhouse and Gravity City journeys. Visual checkpoint: Product Slice 8. Product Slice 9 tests that same Camotion path when A is a filmmaker-supplied photograph rather than a generated still. Observation: Product Slice 9. Product Slice 10 records the Construct prompt requiring visual camera-viewpoint displacement on same-environment journeys, with a deer-forest example. Observation: Product Slice 10. Product Slice 11 records Takes: NEW TAKE appends footage instead of overwriting; the selected Take drives FOOTAGE, preview, and export. Visual checkpoint: Product Slice 11. Product Slice 12 records the first unattended Agent CREATE JOURNEY loop, with jungle minecart and clown funhouse assembled movies. Visual checkpoint: Product Slice 12. Product Slice 13 records four whole-journey camera grammars — POV, FOLLOW, LEAD, MOUNTED — with the same canyon story and four comparison cuts. Visual checkpoint: Product Slice 13. A later evidence spike froze the actual forest destinations A–F, applied Camotion 01.8, generated five Seedance 2.5 legs, and hard-concatenated a review movie outside the product UI. See Forest A→F.

Key findings

• Visual / world continuity ≠ spatial traversability ≠ continuous camera travel
• Continuous locomotion prompting outranks destination matching
• Depth weighting improved motion conditioning on Ghost Library
• Our analysis of Terran's original TunnelVision Photoshop workflow indicates depth-banded exposure compositing rather than a single continuously depth-scaled blur
• Motion-treating the strong depth mask was important in 01.5
• Terminal-at-canonical start exposure worsened Seedance behavior through Krea (01.6). A later Replicate rerun did not reproduce that severe collapse/retreat
• Halving strength removed obvious recursion but weakened geometry authority (01.7)
• 01.8 preserved more central route geometry at still level. Krea later opened backward, then went forward; a Replicate rerun did not reproduce that pronounced reverse-then-forward opening. Those Krea behaviors should not currently be treated as intrinsic to the Camotion variants. 01.8 Route-Preserved Exposure remains the current Camotion baseline, not a promotion over 01.5 as CameraMotionPlan
• Environmental and foreground motion can act as a perceptual witness to camera locomotion. Observation/hypothesis, not a proven mechanism
• On one Seedance prompt-control pair, slowing the locomotion prompt produced a slower traversal; adding walking-embodiment language did not clearly produce footstep-driven motion. One-run observations, not proof
• 01.9 densified 01.8 exposure taps to ≤1 px spacing (2–98 samples). The repeated-object appearance did not clearly go away (MAE vs 01.8 ≈ 0.25, max channel difference 7). Sparse sampling is unlikely to be the dominant cause on this fixture. 01.9 is not promoted. No 01.9 video test
• 01.10 extracted real 01.8 compositor intermediates. Discrete ghost copies are already visible in strong exposure. The multi-layer compositor hypothesis is not supported as the origin of first appearance on this fixture. 01.10 is not promoted. No 01.10 video test
• 01.11 characterized the exposure operator on a synthetic fixture and Ghost Library artifact crops. 01.8 fixed-16 is discrete; 01.9 dense sampling fills the local point streak; structured features still read as transformed copies. Sparse sampling is not the dominant remaining cause. No 01.11 candidate is promoted. No 01.11 video test
• 01.12 tested 01.8 fixed-16 gather at strengths 0.02 / 0.04 / 0.06 / 0.08, pristine vs sigma=1 diagnostic control. No useful pristine operating window: the motion cue becomes useful at about the same point structured duplication becomes objectionable (classification C + D). Exposure-operator tuning is closed. No 01.12 condition is promoted. No 01.12 video test
• Directed A→B requires endpoint fidelity. Canonical spatial authority, Camotion motion conditioning, and generative-video improvisation remain distinct
• Integration Test 01 produced a coherent unattended five-shot loop. A destination object is not enough; the shot needs traversable depth through the transition. Movie-level camera-speed jumps at shot boundaries are currently accidental
• Do not abandon Camotion. Compelling traversal shots occurred with pinned 01.8. The movie-level metric is whether the model shoots the route, not whether A′ is aesthetically clean as a still
• On one controlled Wardrobe A→B Seedance seed (70), Camotion-conditioned endpoints plus the authoritative Cinematographer locomotion prompt were complementary and stronger than either factor alone. Camotion remains motion-state conditioning, not image enhancement. 01.12’s still-image structured-copy limit is not contradicted
• On three already-shootable Wardrobe Loop pairs, scene-aware bounded Camotion strength {.02,.04,.08} was preferred over fixed .08 (human 3–0, independent second 3–0). Strength is a per-canonical Cinematographer choice. Camotion Phase 1 is frozen
• On Wardrobe E→A, the Cinematographer independently judged the direct route NEEDS_INTERMEDIATE, proposed one spatial camera position X, inspected the actual generated X, and rejected both E→X and X→A. Protocol stopped before video. Generated sets are evidence, not assumed success. Research Phase 1 is complete
• Product Slice 1 is the first product UI: Plan states the intended journey; Shoot turns Destinations into positions and Journey into camera travel. Shootability is relational on the route. Fixture-driven; generation is not connected
• Product Slice 2 makes Plan a conversational storyboard: intent enters through Conversation; the storyboard is the planning artifact. Shoot remains production reality. Visual checkpoint: 11 — Product Slice 2
• Product Slice 3 is the first autonomous Director step. Two live Gemini 3.1 Pro plans from the Wardrobe Loop story plus starting frame A organized journeys around traversable thresholds. That is semantic/spatial planning observation, not validated physical shootability. Destination imagery was not generated in that slice. Later work added filmmaker story editing, starting-frame upload, and sequential Derived construction. Observation: 12 — Product Slice 3
• A later fixture-independent Director plan from an uploaded alien salt-flat A grounded on the image's monolith/fissure rather than Wardrobe caption text. First image-conditioned destination construction (FLUX Kontext Pro) strongly preserved world/material continuity and produced substantial forward camera displacement, but did not fully cross the Director-specified threshold. Sequential Derived construction later produced the forest A–F canonicals. Local continuity can remain strong while large visual transformation accumulates, and generated reality can diverge from the original open-loop Director plan. Constructed stills are not shootable evidence. Observation: 13 — Destination construction
• A forest A→F evidence spike on those actual constructed destinations produced a directionally coherent 30-second Camotion + Seedance journey (forest → mouth → tube → crystal → portal → void) without changing Director, Camotion, or product architecture. Destination/world continuity did not automatically produce physical traversal continuity. A→B: convincing forward locomotion. D→E: useful threshold grammar (next space visible through the remaining opening before the previous space disappeared). C→D: early endpoint attraction / weak start authority. E→F: open-void geometry plus radial-forward Camotion read as warp rather than traversal. Human review of the assembled movie was positive. Not a Render Movie product feature. Do not conclude that Camotion failed. Observation: 14 — Forest A→F
• Product Slice 4 is the current Plan | Shoot product UI: untitled project, conversation as turn history, Project panel (story, destination count, auto-generate A, auto-generate all, PLAN), sequential destination construction, and Shoot blocking. The production ladder is to block / blocked / shooting / complete. Visual checkpoint: 15 — Product Slice 4
• Product Slice 5 adds a Project panel Video control so every SHOOT in the current project uses one generator. Pruna remains the development default. The first mid-tier comparison takes (Luma Ray Flash 2 720p, Wan 2.2 First/Last Frame, Seedance 2.0 Fast) are recorded at 16 — Product Slice 5
• Product Slice 6 records destination look-ahead. Construct of N injects N+1's visual description as far-field only; N stays N; the last beat has no look-ahead. Observation: 17 — Product Slice 6
• Product Slice 7 is a Plan | Shoot UI checkpoint on a live untitled ten-destination train journey. Plan shows media facts on the selected still. Shoot uses MOTION Plan / FOOTAGE Generate, with stored Camotion overlay on destination preview. Debug is on by default for now. Visual checkpoint: 18 — Product Slice 7
• Product Slice 8 is the current product: a filmmaking pipeline for continuous first-person journeys, not a prompt wrapper. CREATE JOURNEY structures canonicals; CM describes this shot before the global locomotion baseline; adaptive Camotion conditions endpoints; the video model films the traversal. Nano Banana 2 at 2K is the current strong still configuration. Greenhouse reads as directed plausible continuity; Gravity City keeps journey progression through impossible geometry. AGENT is the next major target. Visual checkpoint: 19 — Product Slice 8
• Product Slice 9 tests Camotion when A is a real photograph: first-person on a causeway, car ahead, strong vanishing point. The still is filmmaker-supplied and authoritative. The take keeps forward travel while weather turns from clear to storm. Not a new Camotion operator. Observation: 20 — Product Slice 9
• Product Slice 10 strengthens canonical Construct so same-environment beats must show a physically advanced camera, not the source composition with new content. Deer forest A–D plus a ~20s first-person take is the example. Not a Director, CM, or Camotion change. Observation: 21 — Product Slice 10
• Product Slice 11 records Takes as first-class footage. NEW TAKE appends TAKE 2, 3… instead of overwriting. The selected Take drives FOOTAGE, preview, playback, and Export Movie. Destination Reshoot is unchanged. Visual checkpoint: 22 — Product Slice 11
• Product Slice 12 records the first JourneyAgent happy path. AGENT CREATE JOURNEY runs unattended through destinations, Motion Plan, NEW TAKE, and assembled Export Movie. Two live runs: jungle minecart and clown funhouse. Visual checkpoint: 23 — Product Slice 12
• Product Slice 13 splits the old first-person default into four whole-journey camera grammars: POV, FOLLOW, LEAD, MOUNTED. Same canyon story, four comparison cuts. Grammar is a Project setting, not a per-shot switch. Visual checkpoint: 24 — Product Slice 13

Chapters

Research

Product