# Scene-aware revision / development experiment 01

Registered before rendering or model outcomes, 2026-09-29.

Question: can a real generated material be bound to an authored moving scene, then survive successive revisions without changing approved geometry, animation or camera? This is a known-state engineering experiment, not a new foundation model or a superiority benchmark.

RFS supplied the constructor/state direction in Computer Week. Our contribution here is the Blender implementation, replayable edit records and measured comparison. Existing UV texturing, procedural rendering and transaction replay are conventional techniques. Library chapters 66, 68 and 72 motivate the conventional baseline, cost accounting and scoped intervention tests.

## Shot and edits

An independently authored stylized courier crosses a gallery while a foreground column occludes it. The camera moves and the courier gestures. 144 frames, 24 fps, 6 seconds, 768×432. Procedural rigid articulation; no cloth simulation or inferred scene state.

One FLUX.2 Klein 4B / 4-bit generation supplies a neutral woven cloth image. Its luminance modulates a coat material in UV coordinates. The RGB image is not a recovered PBR material: illumination/albedo separation is not inferred. A fixed artist-selected tint makes the requested color explicit. Neural generation is confined to material appearance.

Branches: original rust coat; blue coat; blue + moved key light; blue + moved light + column removal; gold coat retaining the light/removal; undo gold back to blue. All branches start from the same saved source and replay a list of typed edits. Natural-language labels are authored here, not parsed by a trained model.

## Comparisons

1. Direct Blender property edits using the same source, assets and requests. Expected equivalent result; this prevents attributing ordinary renderer capabilities to a new algorithm.
2. Source-bound generated material, rendered through the same Blender scene.
3. Stock FLUX.2 Klein image editing on 24 uniformly sampled frames (every sixth frame), 512×288, seed 42, four steps. Input is only the rendered original frame and the same coat instruction. This is a frame-independent image-model diagnostic at 4 fps, NOT a temporal video-model benchmark or equal-information comparison with a scene-aware system.
4. Source + stock edit composited with a supplied coat mask: a stronger conventional protection baseline with access to the same coat mask used in scoring. This does not restore texture correspondence within the coat.

No paid external service, new weights or uploaded production footage. Report all completed samples and failures. GPU workloads run serially under the existing resource guard; preserve 5 GiB available RAM and 12 GiB free disk. Stop on pressure rather than silently relaxing limits.

## Checks / interpretation

* Frozen per-frame geometry, world transforms and camera matrix hashes must match across appearance/light branches; removal may affect only the column's visibility. Check actual evaluated scene state, not only requested parameters.
* Direct Blender and replayed transaction renders at frames 1, 73, 144 should agree in decoded pixels, using identical rendering settings.
* Undo must restore the prior state and rendered frames at 1, 73, 144.
* Source-only negative control and a deliberately shifted protected region check the image metric can distinguish a violation.
* For recoloring, score protected-region RGB MAE and changed-pixel fraction outside a dilated coat mask at 512×288. Report requested-color coverage inside the coat, with a visible-pixel denominator; it is a crude diagnostic, not aesthetic quality. Do not apply background equality to relighting/removal, which legitimately alter shadows and illumination.
* Sampled-frame temporal diagnostics may measure residual changes but must not be called exact motion or identity accuracy. No trained video model is tested in this first local run.
* Track asset generation and render wall time separately. Scene authoring, supplied IDs, UVs and known animation are substantial privileged preparation; no efficiency win over arbitrary-video editing can be claimed.

Success means a reviewable video, actual neural material asset, truthful action log, passing scoped state/undo checks and exposed baseline failures/successes. It does not mean production readiness, automatic scene reconstruction, fair attribution, artist acceptance or advantage over existing 3D workflows.