TunnelVision is an AI filmmaking system for directing continuous first-person journeys through imagined worlds. The filmmaker defines the journey. The Director structures the destinations. The Cinematographer figures out how to physically shoot each traversal. Camotion prepares motion-conditioned endpoint frames. The video model films the traversal. The result is a continuous journey assembled from canonical destinations.
This is not a prompt wrapper around a video model. The current movies work because several layers reinforce each other. Exploratory research has been absorbed into the product architecture; these two journeys show the pieces working as a system.
JOURNEY PROMPT is story intent and journey beats. The filmmaker is not asked to write camera mechanics, locomotion prompting, route geometry, model-specific negatives, or video-generation syntax. CREATE JOURNEY is the product-facing Director action. Agency is DIRECTED or AGENT.
A, B, C, and later letters are pristine canonical destinations. A→B is the traversal between actual adjacent canonicals. A′ and B′ are Camotion-conditioned shooting frames derived from those stills. Do not collapse destination, traversal, and shooting frame into one object. Changing a canonical invalidates that segment's Motion Plan and recomputes it.
SPECIFY. The filmmaker writes a JOURNEY PROMPT. Story first. Camera language stays out of that prompt.
DIRECT. CREATE JOURNEY turns the prompt and the current storyboard into a sequence of destinations. Filmmaker-supplied stills stay authoritative. The Director resolves unspecified decisions; it does not overwrite specified ones.
CONSTRUCT. Canonicals are generated or supplied. Nano Banana 2 at 2K is the current strong still configuration — not a permanent hard dependency. Higher-quality 2K endpoints are also improving downstream video across providers. Canonical continuity is now a major strength of the pipeline.
For each actual adjacent pair, CM inspects the real images and writes a Segment Motion Plan: route, shot direction, pace, subject continuity when useful, Set Consistency, Traversal Confidence, semantic travel geometry (VP and D), and segment-specific video guidance.
The shooting prompt is intentionally split. CM describes this shot's visible physical route first. The global baseline then enforces continuous first-person locomotion and POV invariants. CM names the route that exists — follow the trail, pass between landmarks, emerge onto a shoreline, cross water toward a waterfall — rather than listing hypothetical tunnels or caves the model should not invent. The baseline no longer carries generic route vocabulary.
The baseline principle: continuous forward physical travel; no dissolves, morphs, crossfades, cuts, teleporting, or reversing; spatially contiguous progression; strong natural parallax; unembodied first-person POV; never show the viewer, body, shadow, reflection, or FPS-style objects.
Camotion sits between CM planning and video generation. It is not an agent. Pace still sets base exposure. Adaptive weighting then scales that exposure per pixel:
motion strength ≈ CM pace × depth weighting × destination protection × VP protection
Nearby geometry receives stronger motion. Distant scenery and the arrival / perspective center stay comparatively stable, which helps keep 2K canonical detail. Debug work dirs can show the depth map, destination mask, VP mask, combined weight, and final A′/B′. Destination-aware curved / non-radial VP→D fields are still backlog, not this product path.
It comes from (1) strong canonical continuity, (2) Director journey structure, (3) CM spatial reasoning, (4) segment-specific shot guidance, (5) Camotion endpoint conditioning, (6) capable start/end video generation, and (7) repeated canonical handoffs.
Greenhouse / forest → water → enormous roots → subterranean bioluminescent world. The movie reads as a directed continuous journey rather than stitched AI transitions.
Ordinary city → gravity breaks down → streets and buildings fold vertically → floating blocks → inverted city overhead → intersecting impossible architecture → suspended park. TunnelVision can keep journey progression even when the world itself becomes geometrically impossible. This is not a claim of perfect physical realism. The useful point is that the system directs the impossibility into a coherent traversal.
AGENT should eventually run unattended: journey → canonicals → CM → repair or reshoot when needed → Camotion → footage → evaluate → continue → assembled journey. Planned repair language is SHOOT, RESHOOT_START, RESHOOT_END, RESHOOT_BOTH. When the set is the problem, repair the canonical before trying to solve impossible geography with video prompting. The opposite endpoint may be a visual reference. Traversal Confidence is evidence, not absolute truth — low-confidence pairs can still produce convincing footage. A later policy may shoot high-confidence pairs, shoot medium-confidence first and repair only if footage fails, and repair very low-confidence pairs before shooting. Those thresholds are not final.
Other future work, not this slice: footage evaluator, CM-selected shot duration, filmmaker-adjustable D, destination-aware Camotion, visual look-ahead from future actual canonicals, re-direct around supplied future stills, final journey export, provider bakeoff.