HighVis R&D · Programme Hub · 2026-06-11

Direct the shot in Unreal. Let AI handle the photoreal.

HighVis is a tool-agnostic infrastructure for restyling low-fidelity Unreal renders into photoreal video, while preserving camera, composition, and motion. We're building it three ways in parallel. Each one is a fork: same architecture, different model stack, different infrastructure trade-off. We then compare the forks on the same scenes with the same questions, so the answer still holds when the models change.

Owners

Christopher Jorna · Fivefold

Audience

CoSTAR · Innovate UK · Media Cymru

Forks

Classic · Hybrid · Local

Status

Active · 3-fork bake-off complete · multi-frame checkpoint build started

Recent architectural decisions

Small architectural calls that shape the rest of the build. Documented here so the reasoning sticks.

Programme · 2026-06-11

Two-track split: Switch X vs R&D pipeline

Switch X handles commercial delivery. The ComfyUI pipeline is the R&D track. On a paying project, go straight to Switch X. On R&D, build what Switch X currently lacks. The three things worth building are: (a) better camera and reference control, (b) multi-frame checkpoints, (c) HDR output. Source: Joe Loe call 2026-06-11, logged in state/highvis-bakeoff-2026-06-11.md.

Local · priority build · 2026-06-11

Multi-frame checkpoints fix style fade

The known limitation (F-LC-008): the style anchor is strongest at frame 0 and fades as the clip progresses. The fix (F-LC-009, now in build as V6): generate a fresh restyle frame at each chunk boundary so every chunk starts with a strong style anchor. V6 proved the restyle stage at two keyframes today. Full chunked V2V is the next step. This is the priority build while Seedance API access is pending.

Local · confirmed · 2026-06-11

HDR EXR output confirmed in production

The LTX 2.3 + HDR IC-LoRA pipeline writes a full floating-point EXR sequence alongside the MP4. The EXR retains highlight and shadow detail that the MP4 clips. A +0.2 gain test flagged by CJ confirms the headroom is usable in post. Comparison image pending. This is one of the three confirmed R&D differentiators beyond web-based tools.

Local · limitation found

F-LC-008: The new look fades over time

The AI uses a single reference image (the restyled first frame) to guide the entire video. The look is strongest at the start and fades to about 60% after the first second, then holds. This is a fundamental limitation of using one reference frame for a whole clip. Not fixable by adjusting settings. The multi-frame checkpoint build (V6) is the direct architectural response.

01 · For decision-makers

Exec Summary

A short, fact-locked digest of the programme: what HighVis is, where each fork sits, what's been measured, what's still open.

~3 min read→

02 · For researchers

Living Research Log

The formal academic record. Twelve sections, every claim traced back to a canonical source. Updated as findings land.

~25 min read→

06 · For test data

Test Board

Per-test results board pulling from HighVisRnD/board_data.json. Visual cross-fork comparison surface.

Live test data→