# Spring: two revisions on a previously unused passage

This tests transfer of the guided workflow to a different character, film, aspect ratio and performance. It does not establish a new model capability, an advantage over existing tools, or a globally unseen benchmark.

## Selection and frozen inputs

Spring, 107.75–110.75 seconds: a three-second portrait with a hand-to-mouth gesture, lowering hand, blink and head turn while holding a wooden staff. Source cuts at approximately 107.4583 and 110.9583 seconds fall outside this excerpt. The source is a previously downloaded CC BY4.0 Blender film. Its other passages were used in local monitoring research; this excerpt is outside the ranges recorded in that corpus and had not been used in Afterimage's generative trials. Selection occurred before any new video generation.

The frozen protocol hashes the source, ten guides, two conventional references, prompts, seed and protected decisions. Five guides are at frames0,18,36,54,66, corresponding to0,.75,1.5,2.25,2.75seconds. Review samples include those anchors and unguided frames9,27,45,60,69,71. No outputs are rerolled or repaired between calls. Stage2 receives the exact stage1 output bytes and the original source-derived identity anchors with the predeclared amber grade.

The original1280×536 input is a spatially scaled excerpt of the1920×804 download. Video stays at24fps with72frames; audio is omitted throughout. The returned stage1 video is1920×804 and is carried into stage2 without normalization. Accumulated encoding and synthesized detail are therefore part of continuation, not independently isolated variables.

## Reference construction

One built-in image-generation call made a frozen, character-free forest-floor plate from a source frame. This was separate from Runway's credit balance. Approximate GrabCut masks retain source girl and staff pixels. Source-background ORB/RANSAC registration moves the plate with the shot. The second reference uses the previously specified fixed linear-RGB amber grade. Both complete reference videos were made before the new model outputs.

A preflight mask cut into late-frame hair; source-position seeds were corrected before protocol freeze or video submission. This preparation is authoring, not a result-selected model reroll. These masks are inferred image segmentation, not native geometry or ground truth. Hair edges, blur, optical contact, planar registration and physical illumination remain limitations. Grading is not physically reconstructed relighting.

## Comparison and human review

The conventional path has all72 source masks; Runway receives five guidance images. Both attempt the same visible brief, but this information-access difference prevents an equal-information algorithm comparison. The conventional path is agent-authored, not a timed skilled-compositor completion.

The portable shot-review.html shows source and candidates A/B. Pair order was drawn before generation. No ratings or preferred answer are supplied. The form records creation involvement and prior exposure. It separately asks about transformation, identity, performance timing, staff contact and temporal consistency, then overall preference. “Unclear” and “neither usable” are valid responses. A saved response can be downloaded and is also stored on the local server when available. Nothing is submitted to an external service by that review form.

Method labels are concealed in the initial review view, not cryptographically hidden. Visual clues, earlier tool outputs, downloadable media and the evidence page can reveal methods. A reviewer who has seen those should report exposure. The user helped direct the project; their review is valuable author feedback, but should not automatically be described as independent external judgment. A separate uninvolved reviewer is needed for that claim. No reviewer has been fabricated or contacted.

Independent response count starts at zero. Agent inspection is a separate diagnostic record, never an independent vote. One shot and one seed per stage cannot establish reliability across scenes, preference superiority, labor savings or statistical significance.

## Reproduction and portability

The portable package contains the videos, guide images, protocol, hashes, findings and local review interface. It requires no original source folder to view or review. Re-running prepare_guides.py requires the extracted source frames and local Python/OpenCV/Pillow/numpy/ffmpeg dependencies, which are not all packaged. The raw returned videos are retained because a seed does not guarantee exact API reproduction. Credentials and SDK environment are excluded. Original source files were not changed.

Source: Spring (2019), Blender Animation Studio. © Blender Foundation | cloud.blender.org/spring. CC BY4.0. https://studio.blender.org/projects/spring/pages/about/